Ordinary and Partial Differential Equations
An Introduction to Dynamical Systems
John W. Cain, Ph.D. and Angela M. Reynolds, Ph.D.
Mathematics Textbook Series. Editor: Lon Mitchell
1. Book of Proof by Richard Hammack
2. Linear Algebra by Jim Hefferon
3. Abstract Algebra: Theory and Applications by Thomas Judson
4. Ordinary and Partial Differential Equations by John W. Cain and Angela Reynolds
Department of Mathematics & Applied Mathematics
Virginia Commonwealth University
Richmond, Virginia, 23284
Publication of this edition supported by the Center for Teaching Excellence at vcu
Ordinary and Partial Differential Equations: An Introduction to Dynamical Systems
Edition 1.0
© 2010 by John W. Cain and Angela Reynolds
This work is licensed under the Creative Commons AttributionNonCommercialNo Derivative
Works 3.0 License and is published with the express permission of the authors.
Typeset in 10pt Palladio L with Pazo Math fonts using PDFL
A
T
E
X
Acknowledgements
John W. Cain expresses profound gratitude to his advisor, Dr. David G. Scha
effer, James B. Duke Professor of Mathematics at Duke University. The ﬁrst
ﬁve chapters are based in part upon Professor Schaeffer’s introductory gradu
ate course on ordinary differential equations. The material has been adapted
to accommodate upperlevel undergraduate students, essentially by omitting
technical proofs of the major theorems and including additional examples. Other
major inﬂuences on this book include the excellent texts of Perko [8], Strauss [10],
and Strogatz [11]. In particular, the material presented in the last ﬁve chapters
(including the ordering of the topics) is based heavily on Strauss’ book. On the
other hand, our exposition, examples, and exercises are more “userfriendly”,
making our text more accessible to readers with less background in mathematics.
Dr. Reynolds dedicates her portion of this textbook to her mother, father and
sisters, she thanks them for all their support and love.
Finally, Dr. Cain dedicates his portion of this textbook to his parents Jeanette
and Harry, who he loves more than words can express.
iii
Contents
Acknowledgements iii
Contents iv
1 Introduction 1
1.1 Initial and Boundary Value Problems . . . . . . . . . . . . . . . . . 4
2 Linear, ConstantCoefﬁcient Systems 8
2.1 Homogeneous Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.1 Diagonalizable Matrices. . . . . . . . . . . . . . . . . . . . . 12
2.1.2 Algebraic and Geometric Multiplicities of Eigenvalues. . . . 21
2.1.3 Complex Eigenvalues. . . . . . . . . . . . . . . . . . . . . . . 29
2.1.4 Repeated Eigenvalues and NonDiagonalizable Matrices. . 37
2.2 Phase Portraits and Planar Systems . . . . . . . . . . . . . . . . . . 45
2.3 Stable, Unstable, and Center Subspaces . . . . . . . . . . . . . . . . 57
2.4 Trace and Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.5 Inhomogeneous Systems . . . . . . . . . . . . . . . . . . . . . . . . . 67
3 Nonlinear Systems: Local Theory 78
3.1 Linear Approximations of Functions of Several Variables . . . . . . 81
3.2 Fundamental Existence and Uniqueness Theorem . . . . . . . . . . 84
3.3 Global Existence, Dependence on Initial Conditions . . . . . . . . . 86
3.4 Equilibria and Linearization . . . . . . . . . . . . . . . . . . . . . . . 94
3.5 The HartmanGrobman Theorem . . . . . . . . . . . . . . . . . . . . 98
3.6 The Stable Manifold Theorem . . . . . . . . . . . . . . . . . . . . . . 100
3.7 NonHyperbolic Equilibria and Lyapunov Functions . . . . . . . . 105
4 Periodic, Heteroclinic, and Homoclinic Orbits 122
4.1 Periodic Orbits and the PoincaréBendixon Theorem . . . . . . . . 122
iv
contents v
4.2 Heteroclinic and Homoclinic Orbits . . . . . . . . . . . . . . . . . . 130
5 Bifurcations 140
5.1 Three Basic Bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.2 Dependence of Solutions on Parameters . . . . . . . . . . . . . . . . 148
5.3 AndronovHopf Bifurcations . . . . . . . . . . . . . . . . . . . . . . 151
6 Introduction to Delay Differential Equations 166
6.1 Initial Value Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 168
6.2 Solving ConstantCoefﬁcient Delay Differential Equations . . . . . 169
6.3 Characteristic Equations . . . . . . . . . . . . . . . . . . . . . . . . . 171
6.4 The HutchinsonWright Equation . . . . . . . . . . . . . . . . . . . . 172
7 Introduction to Difference Equations 180
7.1 Basic Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.2 Linear, ConstantCoefﬁcient Difference Equations . . . . . . . . . . 181
7.3 FirstOrder Nonlinear Equations and Stability . . . . . . . . . . . . 191
7.4 Systems of Nonlinear Equations and Stability . . . . . . . . . . . . 195
7.5 PeriodDoubling Bifurcations . . . . . . . . . . . . . . . . . . . . . . 200
7.6 Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.7 How to Control Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . 208
8 Introduction to Partial Differential Equations 218
8.1 Basic Classiﬁcation of Partial Differential Equations . . . . . . . . . 221
8.2 Solutions of Partial Differential Equations . . . . . . . . . . . . . . . 227
8.3 Initial Conditions and Boundary Conditions . . . . . . . . . . . . . 228
8.4 Visualizing Solutions of Partial Differential Equations . . . . . . . . 233
9 Linear, FirstOrder Partial Differential Equations 236
9.1 Derivation and Solution of the Transport Equation . . . . . . . . . 239
9.2 Method of Characteristics: More Examples . . . . . . . . . . . . . . 241
10 The Heat and Wave Equations on an Unbounded Domain 250
10.1 Derivation of the Heat and Wave Equations . . . . . . . . . . . . . . 250
10.2 Cauchy Problem for the Wave Equation . . . . . . . . . . . . . . . . 255
10.3 Cauchy Problem for the Heat Equation . . . . . . . . . . . . . . . . 265
10.4 WellPosedness and the Heat Equation . . . . . . . . . . . . . . . . 276
10.5 Inhomogeneous Equations and Duhamel’s Principle . . . . . . . . 284
vi
11 InitialBoundary Value Problems 297
11.1 Heat and Wave Equations on a HalfLine . . . . . . . . . . . . . . . 297
11.2 Separation of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 306
11.2.1 Wave Equation, Dirichlet Problem. . . . . . . . . . . . . . . . 307
11.2.2 Heat Equation, Dirichlet Problem. . . . . . . . . . . . . . . . 313
11.2.3 Wave Equation, Neumann Problem. . . . . . . . . . . . . . . 318
11.2.4 Heat Equation, Neumann Problem. . . . . . . . . . . . . . . 324
11.2.5 Mixed Boundary Conditions: An Example. . . . . . . . . . . 324
12 Introduction to Fourier Series 330
12.1 Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
12.1.1 Fourier sine series. . . . . . . . . . . . . . . . . . . . . . . . . 332
12.1.2 Fourier cosine series. . . . . . . . . . . . . . . . . . . . . . . . 337
12.1.3 Fourier series. . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
12.2 Convergence of Fourier Series . . . . . . . . . . . . . . . . . . . . . . 344
12.2.1 Norms, distances, inner products, and convergence. . . . . 347
12.2.2 Convergence theorems. . . . . . . . . . . . . . . . . . . . . . 359
13 The Laplace and Poisson Equations 367
13.1 Dirchlet and Neumann Problems . . . . . . . . . . . . . . . . . . . . 370
13.2 Wellposedness and the Maximum Principle . . . . . . . . . . . . . 372
13.3 Translation and Rotation Invariance . . . . . . . . . . . . . . . . . . 375
13.4 Laplace’s Equation on Bounded Domains . . . . . . . . . . . . . . . 383
13.4.1 Dirichlet problem on a rectangle. . . . . . . . . . . . . . . . . 383
13.4.2 Dirichlet problem on a disc. . . . . . . . . . . . . . . . . . . . 390
Guide to Commonly Used Notation 404
References 406
Index 407
CHAPTER 1
Introduction
T
he mathematical subdiscipline of differential equations and dynamical systems
is foundational in the study of applied mathematics. Differential equations
arise in a variety of contexts, some purely theoretical and some of practical
interest. As you read this textbook, you will ﬁnd that the qualitative and
quantitative study of differential equations incorporates an elegant blend of linear
algebra and advanced calculus. For this reason, it is expected that the reader has
already completed courses in (i) linear algebra; (ii) multivariable calculus; and
(iii) introductory differential equations. Familiarity with the following topics is
especially desirable:
From basic differential equations: separable differential equations and separa
tion of variables; and solving linear, constantcoefﬁcient differential equations
using characteristic equations.
From linear algebra: solving systems of m algebraic equations with n un
knowns; matrix inversion; linear independence; and eigenvalues/eigenvectors.
From multivariable calculus: parametrized curves; partial derivatives and
gradients; and approximating a surface using a tangent plane.
Some of these topics will be reviewed as we encounter them later—in this
chapter, we will recall a few basic notions from an introductory course in
differential equations. Readers are encouraged to supplement this book with the
excellent textbooks of Hubbard and West [5], Meiss [7], Perko [8], Strauss [10],
and Strogatz [11].
Question: Why study differential equations?
1
2
Answer: When scientists attempt to mathematically model various natural
phenomena, they often invoke physical “laws” or biological “principles” which
govern the rates of change of certain quantities of interest. Hence, the equations
in mathematical models tend to include derivatives. For example, suppose
that a hot cup of coffee is placed in a room of constant ambient temperature α.
Newton’s Law of Cooling states that the rate of change of the coffee temperature
T(t) is proportional to the difference between the coffee’s temperature and the
room temperature. Mathematically, this can be expressed as
dT
dt
= k(T − α),
where k is a proportionality constant.
Solution techniques for differential equations (des) depend in part upon how
many independent variables and dependent variables the system has.
Example 1.0.1. One independent variable and one independent variable. In
writing the equation
d
2
y
dx
2
+cos(xy) = 3,
it is understood that y is the dependent variable and x is the independent
variable.
When a differential equation involves a single independent variable, we refer
to the equation as an ordinary differential equation (ode).
Example 1.0.2. If there are several dependent variables and a single independent
variable, we might have equations such as
dy
dx
= x
2
y −xy
2
+ z,
dz
dx
= z −y cos x.
This is a system of two odes, and it is understood that x is the independent
variable.
Example 1.0.3. One dependent variable, several independent variables. Consider
the de
∂u
∂t
=
∂
2
u
∂x
2
+
∂
2
u
∂y
2
.
This equation involves three independent variables (x, y, and t) and one depen
dent variable, u. This is an example of a partial differential equation (pde). If there
are several independent variables and several dependent variables, one may have
systems of pdes.
i ntroducti on 3
Although these concepts are probably familiar to the reader, we give a more
exact deﬁnition for what we mean by ode. Suppose that x and y are independent
and dependent variables, respectively, and let y
(k)
(x) denote the kth derivative
of y with respect to x. (If k ≤ 3, we will use primes.)
Deﬁnition 1.0.4. Any equation of the form F(x, y, y
/
, y
//
, . . . , y
(n)
) = 0 is called
an ordinary differential equation. If y
(n)
is the highest derivative appearing in the
equation, we say that the ode is of order n.
Example 1.0.5.
_
d
3
y
dx
3
_
2
−(cos x)
dy
dx
= y
d
2
y
dx
2
can be written as (y
///
)
2
−yy
//
−(cos x)y
/
= 0, so using the notation in the above
Deﬁnition, we would have F(x, y, y
/
, y
//
, y
///
) = (y
///
)
2
−yy
//
−(cos x)y
/
. This is a
thirdorder ode.
Deﬁnition 1.0.6. A solution of the ode F(x, y, y
/
, y
//
, . . . , y
(n)
) = 0 on an interval
I is any function y(x) which is ntimes differentiable and satisﬁes the equation
on I.
Example 1.0.7. For any choice of constant A, the function
y(x) =
Ae
x
1 + Ae
x
is a solution of the ﬁrstorder ode y
/
= y −y
2
for all real x. To see why, we use
the quotient rule to calculate
y
/
=
Ae
x
(1 + Ae
x
) −(Ae
x
)
2
(1 + Ae
x
)
2
=
Ae
x
(1 + Ae
x
)
2
.
By comparison, we calculate that
y −y
2
=
Ae
x
(1 + Ae
x
)
−
(Ae
x
)
2
(1 + Ae
x
)
2
=
Ae
x
(1 + Ae
x
)
2
.
Therefore, y
/
= y −y
2
, as claimed.
The deﬁnition of a solution of an ode is easily extended to systems of odes
(see below). In what follows, we will focus solely on systems of ﬁrstorder odes.
This may seem overly restrictive, until we make the following observation.
4 i ni ti al and boundary value problems
Observation. Any nthorder ode can be written as a system of n ﬁrstorder
odes. The process of doing so is straightforward, as illustrated in the following
example:
Example 1.0.8. Consider the secondorder ode y
//
+ (cos x)y
/
+y
2
= e
x
. To avoid
using second derivatives, we introduce a new dependent variable z = y
/
so that
z
/
= y
//
. Our ode can be rewritten as z
/
+ (cos x)z + y
2
= e
x
. Thus, we have
obtained a system of two ﬁrstorder odes:
dy
dx
= z,
dz
dx
= −(cos x)z −y
2
+e
x
.
A solution of the above system of odes on an open interval I is any vector
of differentiable functions [y(x), z(x)] which simultaneously satisfy both odes
when x ∈ I.
Example 1.0.9. Consider the system
dy
dt
= z,
dz
dt
= −y.
We claim that for any choices of constants C
1
and C
2
,
_
y(t)
z(t)
_
=
_
C
1
cos t + C
2
sin t
−C
1
sin t + C
2
cos t
_
is a solution of the system. To verify this, assume that y and z have this form.
Differentiation reveals that y
/
= −C
1
sin t + C
2
cos t and z
/
= −C
1
cos t −C
2
sin t.
Thus, y
/
= z and z
/
= −y, as required.
1.1. Initial and Boundary Value Problems
In the previous example, the solution of the system of odes contains arbitrary
constants C
1
and C
2
. Therefore, the system has inﬁnitely many solutions. In
practice, one often has additional information about the underlying system,
allowing us to select a particular solution of practical interest. For example,
suppose that a cup of coffee is cooling off and obeys Newton’s Law of Cooling.
In order to predict the coffee’s temperature at future times, we would need to
specify the temperature of the coffee at some reference time (usually considered
to be the “initial” time). By specifying auxiliary conditions that solutions of an
i ntroducti on 5
ode must satisfy, we may be able to single out a particular solution. There are
two usual ways of specifying auxiliary conditions.
Initial conditions. Suppose F(x, y, y
/
, y
//
, . . . , y
(n)
) = 0 is an nth order ode
which has a solution on an open interval I containing x = x
0
. Recall from
your course on basic differential equations that, under reasonable assumptions,
we would expect the general solution of this ode to contain n arbitrary constants.
One way to eliminate these constants and single out one particular solution is to
specify n initial conditions. To do so, we may specify values for
y(x
0
), y
/
(x
0
), y
//
(x
0
), . . . y
(n−1)
(x
0
).
We regard x
0
as representing some “initial time”. An ode together with its initial
conditions (ics) forms an initial value problem (ivp). Usually, initial conditions
will be speciﬁed at x
0
= 0.
Example 1.1.1. Consider the secondorder ode y
//
(x) + y(x) = 0. You can check
that the general solution is y(x) = C
1
cos x + C
2
sin(x), where C
1
and C
2
are
arbitrary constants. To single out a particular solution, we would need to specify
two initial conditions. For example, if we require that y(0) = 1 and y
/
(0) = 0, we
ﬁnd that C
1
= 1 and C
2
= 0. Hence, we obtain a particular solution y(x) = cos x.
If we have a system of n ﬁrstorder odes, we will specify one initial condition
for each independent variable. If the dependent variables are
y
1
(x), y
2
(x), . . . y
n
(x),
we typically specify the values of
y
1
(0), y
2
(0), . . . , y
n
(0).
Boundary conditions. Instead of specifying requirements that y and its deriva
tives must satisfy at one particular value of the independent variable x, we could
instead impose requirements on y and its derivatives at different x values. The
result is called a boundary value problem (bvp).
Example 1.1.2. Consider the boundary value problem y
//
+ y = 0 with boundary
conditions y(0) = 1 and y(π/2) = 0. The general solution of the ode is
y(x) = C
1
cos x + C
2
sin x. Using the ﬁrst boundary condition, we ﬁnd that
6 i ni ti al and boundary value problems
C
1
= 1. Since y
/
(x) = −C
1
sin x + C
2
cos x, the second boundary condition tells
us that −C
1
= 0. Notice that the two boundary conditions produce conﬂicting
requirements on C
1
. Consequently, the bvp has no solutions.
As the previous example suggests, boundary value problems can be a tricky
matter. In the ode portion of this text, we consider only initial value problems.
Exercises
1. Write the equation of the line that passes through the points (−1, 2, 3) and
(4, 0, −1) in R
3
, threedimensional Euclidean space.
2. Find the general solution of the differential equation
d
3
y
dx
3
+2
d
2
y
dx
2
+5
dy
dx
= 0.
3. Find the general solution of the differential equation
d
2
y
dx
2
+6
dy
dx
+9y = 0.
4. Solve the ivp
y
//
−3y
/
+2y = 0, y(0) = 1, y
/
(0) = 1.
5. Solve (if possible) the bvp
y
//
−3y
/
+2y = 0, y(0) = 0, y(1) = e.
6. Solve the ivp
y
(4)
−y
//
= 0,
y(0) = 1, y
/
(0) = 0,
y
//
(0) = −1, y
///
(0) = 0.
7. Solve the differential equation
dy
dx
= (y +2)(y +1).
i ntroducti on 7
8. Solve the ivp
dy
dx
= e
y
sin x, y(0) = 0.
9. Find the equations of the planes tangent to the surface
z = f (x, y) = x
2
−2x + y
2
−2y +2
at the points (x, y, z) = (1, 1, 0) and (x, y, z) = (0, 2, 2).
10. Find the eigenvalues of the matrix
A =
_
1 4
4 1
_
and, for each eigenvalue, ﬁnd a corresponding eigenvector.
11. Find the eigenvalues of the matrix
A =
_
¸
_
1 3 −1
0 3 0
0 1 2
_
¸
_
and, for each eigenvalue, ﬁnd a corresponding eigenvector.
12. Write the following differential equations as systems of ﬁrstorder odes:
y
//
−5y
/
+6y = 0
−y
//
−2y
/
= 7 cos(y
/
)
y
(4)
−y
//
+8y
/
+ y
2
= e
x
.
CHAPTER 2
Linear, ConstantCoeﬃcient Systems
T
here are few classes of odes for which exact, analytical solutions can be
obtained by hand. However, for many systems which cannot be solved
explicitly, we may approximate the dynamics by using simpler systems of odes
which can be solved exactly. This often allows us to extract valuable qualitative
information about complicated dynamical systems. We now introduce tech
niques for systematically solving linear systems of ﬁrstorder odes with constant
coefﬁcients.
Notation. Because we will be working with vectors of dependent variables, we
should establish (or recall) some commonly used notation. We denote the set of
real numbers by R. We let R
n
denote the set of all vectors with n components,
each of which is a real number. Usually, vectors will be denoted by bold letters
such as x, y, and we will use capital letters such as A to denote n n matrices
of real numbers. Generally, we shall not distinguish between row vectors and
column vectors, as our intentions will usually be clear from the context. For
example, if we write the product xA, then x should be treated as a row vector,
whereas if we write Ax, then x is understood to be a column vector. If we write
x(t), we mean a vector of functions, each of which depends on a variable t. In
such cases, the vector x(0) would be a constant vector in which each component
function has been evaluated at t = 0. Moreover, the vector x
/
(t) is the vector
consisting of the derivatives of the functions which form the components of x(t).
Systems with constant coefﬁcients. Suppose that y
1
, y
2
, . . . y
n
are variables
which depend on a single variable t. The general form of a linear, constant
8
li near, constant coeffi ci ent systems 9
coefﬁcient system of ﬁrstorder odes is as follows:
dy
1
dt
= a
11
y
1
(t) + a
12
y
2
(t) + · · · + a
1n
y
n
(t) + f
1
(t)
dy
2
dt
= a
21
y
1
(t) + a
22
y
2
(t) + · · · + a
2n
y
n
(t) + f
2
(t)
.
.
.
dy
n
dt
= a
n1
y
1
(t) + a
n2
y
2
(t) + · · · + a
nn
y
n
(t) + f
n
(t).
(2.1)
Here, each a
ij
is a constant (1 ≤ i, j ≤ n), and f
i
(t) (i = 1, 2, . . . n) are functions
of t only.
Example 2.0.3. Soon, we will learn how to solve the linear, constantcoefﬁcient
system
dy
1
dt
= 3y
1
−2y
2
+cos t
dy
2
dt
= 10y
2
−t
2
+6.
(2.2)
The system (2.1) can be written more compactly if we introduce matrix/vector
notation. Suppressing the dependence on t for notational convenience, routine
matrix multiplication will verify that
_
¸
¸
¸
¸
¸
_
y
/
1
y
/
2
.
.
.
y
/
n
_
¸
¸
¸
¸
¸
_
=
_
¸
¸
¸
¸
¸
_
a
11
a
12
. . . a
1n
a
21
a
22
. . . a
2n
.
.
.
.
.
.
.
.
.
.
.
.
a
n1
a
n2
. . . a
nn
_
¸
¸
¸
¸
¸
_
_
¸
¸
¸
¸
¸
_
y
1
y
2
.
.
.
y
n
_
¸
¸
¸
¸
¸
_
+
_
¸
¸
¸
¸
¸
_
f
1
f
2
.
.
.
f
n
_
¸
¸
¸
¸
¸
_
(2.3)
is equivalent to the system (2.1). Furthermore, if we deﬁne
y =
_
¸
¸
¸
¸
¸
_
y
1
y
2
.
.
.
y
n
_
¸
¸
¸
¸
¸
_
y
/
=
_
¸
¸
¸
¸
¸
_
y
/
1
y
/
2
.
.
.
y
/
n
_
¸
¸
¸
¸
¸
_
f =
_
¸
¸
¸
¸
¸
_
f
1
f
2
.
.
.
f
n
_
¸
¸
¸
¸
¸
_
(2.4)
and let A denote the matrix of coefﬁcients in Equation (2.3), then the entire
system takes the form
y
/
= Ay + f. (2.5)
10 homogeneous systems
Deﬁnition 2.0.4. The system of odes in Equation (2.5) is called homogeneous if
the vector f is the zero vector. Otherwise, the system is inhomogeneous.
Example 2.0.5. The system
dy
1
dt
= 5y
1
−y
2
dy
2
dt
= −y
1
+8y
2
is homogeneous, whereas the system (2.2) above is inhomogeneous. Notice that
we may write (2.2) in matrix notation as
_
y
/
1
y
/
2
_
=
_
3 −2
0 10
_ _
y
1
y
2
_
+
_
cos t
−t
2
+6
_
.
Road map of things to come. In the next subsection, we will learn how to solve
homogeneous constantcoefﬁcient systems y
/
= Ay. This will require lots of linear
algebra! Next, we will learn how to solve inhomogeneous systems of the form (2.5).
Finally, the next several chapters will be devoted to understanding how to
qualitatively analyze solutions of nonlinear systems (which generally cannot be
solved by hand). This will be accomplished by approximating nonlinear systems
with linear ones of the form (2.5).
2.1. Homogeneous Systems
In order to motivate the techniques we will use to solve homogeneous systems
y
/
= Ay, we draw an analogy with a simple, ﬁrstorder initial value problem.
Motivating example. Consider the ode
dy
dt
= ay where a is a constant, and
suppose we have an initial condition y(0) = y
0
. This initial value problem is
easy to solve using techniques from your ﬁrst course in differential equations.
The solution is y(t) = e
at
y
0
. Notice that the solution involves an exponential
function.
Question: Can we extend this example to homogeneous, constantcoefﬁcient
systems of equations?
More speciﬁcally, consider the system y
/
= Ay, where y = y(t) is a vector of
length n and A is an n n constant matrix. Suppose we form an initial value
li near, constant coeffi ci ent systems 11
problem by specifying a vector of initial conditions; i.e., we assume that
y
0
=
_
¸
¸
¸
¸
¸
_
y
1
(0)
y
2
(0)
.
.
.
y
n
(0)
_
¸
¸
¸
¸
¸
_
is given. The resulting ivp has the form y
/
= Ay, y(0) = y
0
, which resembles the
form of the ivp in the motivating example. By analogy with the above system,
can we say that the solution of our new ivp is given by y(t) = e
tA
y
0
? If so, we
would need some way of assigning meaning to the object e
tA
where A is a matrix.
Certainly e
tA
would need to be an n n matrix as well, because both y(t) and y
0
are vectors in R
n
. Fortunately, for square matrices such as A, there is a natural
way to assign meaning to e
A
.
Deﬁnition 2.1.1. Suppose A is an n n constant matrix. The matrix exponential
e
A
is deﬁned in terms of the Maclaurin series expansion of the usual exponential
function. That is,
e
A
=
∞
∑
k=0
1
k!
A
k
, (2.6)
which is a sum involving powers of the matrix A.
We make several remarks about this deﬁnition:
For positive integers k, recall that k! is read “k factorial” and is deﬁned
to be the product of the ﬁrst k natural numbers: k(k −1)(k −2) · · · (3)(2)(1).
Additionally, we deﬁne 0! = 1.
If A is a matrix, then A
0
= I, the identity matrix. (This is analogous to the
fact that a
0
= 1 for scalars a.)
If t is a scalar, then tA is matrix, so
e
tA
=
∞
∑
k=0
1
k!
(tA)
k
=
∞
∑
k=0
t
k
k!
A
k
.
Example 2.1.2. Suppose A is the 2 2 matrix consisting entirely of zeros. Then
A
0
= I, and for each k ≥ 1 we have
A
k
=
_
0 0
0 0
_
.
12 homogeneous systems
Therefore, e
tA
= I because the only nonzero term in the series expansion (2.6) is
the one corresponding to k = 0. Notice the parallel between the fact that e
0
= 1
for scalars and the fact that the exponential of the zero matrix is the identity
matrix.
As we shall soon see, every constant, square matrix A has a matrix exponential
e
A
. Moreover, we will ﬁnd that the solution of the initial value problem y
/
=
Ay, y(0) = y
0
really is y(t) = e
tA
y
0
, just as we would hope. Hence, our
main immediate challenge is to devise a procedure for calculating the matrix
exponential e
tA
for an arbitrary square matrix A. Because e
tA
is deﬁned by its
series representation, we need a method for computing powers of A, which will
require us to review some facts from linear algebra.
2.1.1 Diagonalizable Matrices. Before we recall what it means for a matrix to be
diagonalizable, we consider a very simple class of matrices: the diagonal matrices.
A matrix is diagonal if all of the entries off of the main diagonal are 0. In other
words, a
ij
= 0 whenever i ,= j. When writing diagonal matrices, we typically do
not write the entries which lie off the main diagonal, as it is understood that all
of these entries are 0.
Diagonal matrices. Suppose D is an n n diagonal matrix
D =
_
¸
¸
¸
¸
¸
_
d
1
d
2
.
.
.
d
n
_
¸
¸
¸
¸
¸
_
,
which we will sometimes denote by D = diag{d
1
, d
2
, . . . d
n
}. Straightforward
induction veriﬁes that powers of D are given by
D
k
=
_
¸
¸
¸
¸
¸
_
d
k
1
d
k
2
.
.
.
d
k
n
_
¸
¸
¸
¸
¸
_
= diag{d
k
1
, d
k
2
, . . . d
k
n
}.
li near, constant coeffi ci ent systems 13
According to the series representation for the matrix exponential, we have
e
tD
=
∞
∑
k=0
t
k
k!
D
k
=
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
∞
∑
k=0
t
k
k!
d
k
1
∞
∑
k=0
t
k
k!
d
k
2
.
.
.
∞
∑
k=0
t
k
k!
d
k
n
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
.
The entries in this matrix are simply the Maclaurin series representations for
the functions e
d
1
t
, e
d
2
t
, . . . e
d
n
t
. Therefore, we have shown that for the diagonal
matrix D,
e
tD
=
_
¸
¸
¸
¸
¸
_
e
d
1
t
e
d
2
t
.
.
.
e
d
n
t
_
¸
¸
¸
¸
¸
_
.
Example 2.1.3. Consider the initial value problem
dx
dt
= 6x,
dy
dt
= −3y, x(0) = 2,
and y(0) = 1. These odes are uncoupled in the sense that they are effectively
independent of one another—changes in x(t) have no impact on y(t) and vice
versa. In matrix notation, this system takes the form
_
x
/
y
/
_
=
_
6 0
0 −3
_ _
x
y
_
,
_
x(0)
y(0)
_
=
_
2
1
_
.
Since the coefﬁcient matrix D is diagonal, we immediately have
e
tD
=
_
e
6t
0
0 e
−3t
_
.
Multiplying the matrix exponential by the vector of initial conditions should give
the solution of the initial value problem:
_
x(t)
y(t)
_
= e
tD
_
x(0)
y(0)
_
=
_
e
6t
0
0 e
−3t
_ _
2
1
_
=
_
2e
6t
e
−3t
_
.
You can verify that x(t) = 2e
6t
, y(t) = e
−3t
is, indeed, a solution of this ivp.
14 homogeneous systems
Obviously the diagonal matrices form a very narrow class of matrices, and
most of the systems of odes we wish to consider will be coupled. After the
diagonal matrices, the next easiest class of matrices to exponentiate are the
diagonalizable matrices. Before deﬁning what this means, we recall one other
deﬁnition.
Deﬁnition 2.1.4. Two n n matrices A and B are called similar if there exists an
invertible matrix P such that A = PBP
−1
.
Note that every matrix A is similar to itself (just take P = I, the identity
matrix). It is also easy to show that if A is similar to B, then B is similar to A.
Finally, if A is similar to B and B is similar to C, then A is similar to C. In other
words, similarity of matrices forms an equivalence relation, and matrices can be
partitioned into equivalence classes according to similarity.
Deﬁnition 2.1.5. A square matrix A is called diagonalizable if it is similar to a
diagonal matrix.
Above, we implied that diagonalizable matrices are “nice” because it is straight
forward to exponentiate such matrices. To see why, suppose that A is diago
nalizable. Then there exists a diagonal matrix D and an invertible matrix P
such that A = PDP
−1
. It follows that A
2
= (PDP
−1
)(PDP
−1
). Since matrix
multiplication is associative (although NOT commutative), we may regroup the
terms in this product as long as we preserve their ordering:
A
2
= PD(P
−1
P)DP
−1
= PDIDP
−1
= PD
2
P
−1
.
The same idea can be extended to higher powers of A:
A
n
= (PDP
−1
)(PDP
−1
) · · · (PDP
−1
)
. ¸¸ .
n times
= PD(P
−1
P)D(P
−1
P) · · · (P
−1
P)DP
−1
,
which simpliﬁes to PD
n
P
−1
. Recalling the series expansion for e
tA
, we have
e
tA
=
∞
∑
k=0
t
k
k!
A
k
= lim
n→∞
n
∑
k=0
t
k
k!
A
k
.
li near, constant coeffi ci ent systems 15
Replacing A
k
with PD
k
P
−1
, the series becomes
e
tA
= lim
n→∞
n
∑
k=0
t
k
k!
PD
k
P
−1
.
Since each term in the (ﬁnite) sum is premultiplied by P and postmultiplied by
P
−1
, we may use associativity of matrix multiplication to write
e
tA
= P
_
lim
n→∞
n
∑
k=0
t
k
k!
D
k
_
P
−1
= P
_
∞
∑
k=0
t
k
k!
D
k
_
P
−1
= Pe
tD
P
−1
. (2.7)
Nice consequence: Formula (2.7) suggests a convenient procedure for expo
nentiating a diagonalizable matrix A. First, ﬁnd a diagonal matrix D and an
invertible matrix P such that A = PDP
−1
. After decomposing A in this way, it
follows that e
tA
= Pe
tD
P
−1
, and e
tD
is easy to compute. Of course, the chal
lenging part is ﬁnding the matrices D and P, a task that we shall undertake
shortly.
Before giving a general procedure for exponentiating diagonalizable matrices,
we address a more basic question, namely, “Which matrices are diagonalizable?”
This question, which we answer in Theorem 2.1.7 below, will require us to recall
a deﬁnition from basic linear algebra:
Deﬁnition 2.1.6. Suppose A is an n n matrix. A scalar λ is called an eigenvalue
for A if there exists a nonzero vector v such that Av = λv. Any nonzero vector
v satisfying this equality is called an eigenvector corresponding to the eigenvalue
λ.
Theorem 2.1.7. An n n matrix A is diagonalizable if and only if there exists a
set of n linearly independent eigenvectors for A.
Proof. See your linear algebra textbook. One direction of the proof is fairly
straightforward: If there are n linearly independent eigenvectors, form the matrix
P by letting the columns of P be the eigenvectors. Then, show that A = PDP
−1
,
where D is a diagonal matrix whose diagonal entries are the eigenvalues.
It is important to note that eigenvalues need not be real numbers, and that
eigenvectors may contain complex entries. In what follows, we focus on matrices
which can be diagonalized over the real number ﬁeld. That is, whenever we
write A = PDP
−1
where D is diagonal, we will insist that both D and P have
16 homogeneous systems
u
Au
v
Av
Figure 2.1. The vector u is an eigenvector for A because the vectors u
and Au span the same line. However, v is not an eigenvector because
the lines spanned by v and Av are different.
real entries. Later, when we study matrices with complex eigenvalues, we will
adopt a very different approach to diagonalization.
Geometrically, nonzero real eigenvalues λ and their corresponding eigen
vectors are easily understood. If λ > 0 and v is a corresponding eigenvector,
then the product Av = λv effectively stretches v without altering its direction.
Likewise, if λ < 0, then Av = λv stretches the vector v by a factor λ and
reverses its orientation. Importantly, the lines in R
n
spanned by eigenvectors are
unaffected by multiplication by the matrix A. This is what makes eigenvectors very
special: whereas most vectors are both stretched and rotated via matrixvector
multiplication, eigenvectors have their directions preserved.
Example 2.1.8. If
A =
_
−1 0
−2 2
_
, u =
_
0
1
_
, and v =
_
1
0
_
,
then
Au =
_
0
2
_
and Av =
_
−1
−2
_
.
Observe that Au = 2u, which means that u is an eigenvector and λ = 2 is its
corresponding eigenvalue. Multiplication by A stretches u by a factor of 2 and
preserves its orientation. On the other hand, v is not an eigenvector, because Av
is not simply a scalar multiple of v (see Figure 2.1).
li near, constant coeffi ci ent systems 17
Calculation of eigenvalues is typically done as follows. If we seek a nonzero
vector v satisfying a relationship Av = λv for some scalar λ, then Av −λIv = 0.
Equivalently, (A −λI)v = 0. Since v ,= 0, it must be the case that the matrix
A −λI is singular which, in turn, implies that its determinant is 0. Therefore,
the eigenvalues must satisfy the equation det(A − λI) = 0. Once we have
the eigenvalues, ﬁnding eigenvectors is straightforward, and then the proof of
Theorem 2.1.7 suggests a procedure for diagonalizing A (see below).
Deﬁnition 2.1.9. Given a square matrix A, the equation det(A − λI) = 0 is
called the characteristic equation of A. The expression det(A−λI) is a polynomial
in the variable λ, and is called the characteristic polynomial of A.
How to diagonalize a (diagonalizable) matrix. We will illustrate this method
via an example, using
A =
_
1 −1
2 4
_
.
Step 1: Find the eigenvalues of A by solving the characteristic equation. Since
A −λI =
_
1 −1
2 4
_
−
_
λ 0
0 λ
_
=
_
1 −λ −1
2 4 −λ
_
,
the characteristic polynomial p(λ) is given by
p(λ) = det(A −λI) = (1 −λ)(4 −λ) +2 = (λ −2)(λ −3).
The roots of this polynomial, λ = 2 and λ = 3, are the eigenvalues of A.
Step 2: For each eigenvalue, ﬁnd a maximal linearly independent set of eigen
vectors. For a given eigenvalue λ, if you convert the matrix (A −λI) to reduced
rowechelon form, the number of allzero rows tells you how many linearly inde
pendent eigenvectors you must produce. Let’s start with λ = 2. Eigenvectors
satisfy the equation (A −λI)v = 0 which, in this case, means
_
−1 −1
2 2
_ _
v
1
v
2
_
=
_
0
0
_
.
18 homogeneous systems
Rowreducing this linear, homogeneous system yields
_
1 1
0 0
_ _
v
1
v
2
_
=
_
0
0
_
.
Remark: You should always obtain at least one row of zeros when rowreducing
A −λI. After all, we knew in advance that det(A −λI) = 0. Since our reduced
matrix has one row of zeros, we have one free variable: let us choose v
2
. Ex
panding the system, we have v
1
+ v
2
= 0, which means v
1
= −v
2
. The set of
solutions is therefore
_
v
1
v
2
_
=
_
−v
2
v
2
_
= v
2
_
−1
1
_
,
where v
2
is any real constant. It follows that
_
−1
1
_
is an eigenvector corresponding to λ = 2 (as is any nonzero multiple of this
vector). For λ = 3, the same procedure produces another eigenvector. In this
case, we must solve (A −3I)v = 0, which gives
_
−2 −1
2 1
_ _
v
1
v
2
_
=
_
0
0
_
.
Rowreducing as before,
_
1
1
2
0 0
_ _
v
1
v
2
_
=
_
0
0
_
.
Expanding the last equation reveals that v
1
+
1
2
v
2
= 0. Treating v
2
as the free
variable, we write v
1
= −
1
2
v
2
so that
_
v
1
v
2
_
=
_
−
1
2
v
2
v
2
_
= v
2
_
−
1
2
1
_
.
li near, constant coeffi ci ent systems 19
Thus, any nonzero scalar multiple of
_
−
1
2
1
_
is an eigenvector corresponding to λ = 3. To facilitate computation by hand, let
us choose an eigenvector with integer entries:
_
−1
2
_
.
Step 3: Work with a basis of eigenvectors instead of the standard basis vectors.
To do so, we arrange the eigenvectors as columns in a matrix P. (We usually order
the eigenvectors according to which eigenvalue they correspond to, starting with
the lowest eigenvalues—feel free to choose whatever order you wish.) Speciﬁcally,
let
P =
_
−1 −1
1 2
_
,
the ﬁrst column being the eigenvector corresponding to λ = 2 and the second
column being the eigenvector corresponding to λ = 3. Next, form the diagonal
matrix D whose entries are the eigenvalues:
D =
_
2 0
0 3
_
.
Note that the ordering of the eigenvalues in the matrix D should correspond to
the ordering of the columns in P. Next, the reader can verify that
P
−1
=
_
−2 −1
1 1
_
.
Moreover, straightforward matrix multiplication reveals that
A = PDP
−1
. (2.8)
The process of writing A in the form (2.8) is called diagonalization. After diagonal
izing A, the process of calculating the matrix exponential e
tA
is straightforward.
20 homogeneous systems
Recalling that e
tA
= Pe
tD
P
−1
, we may now calculate
e
tA
=
_
−1 −1
1 2
_ _
e
2t
0
0 e
3t
_ _
−2 −1
1 1
_
=
_
2e
2t
−e
3t
e
2t
−e
3t
−2e
2t
+2e
3t
−e
2t
+2e
3t
_
.
Now that we have substantially widened the class of matrices that we know
how to exponentiate, we formally state a theorem regarding how the matrix
exponential can be used to solve homogeneous systems of constantcoefﬁcient
odes.
Theorem 2.1.10. Suppose that y
/
= Ay is a system of constant coefﬁcient odes
and let y
0
= y(0) be a vector of initial conditions. Then this initial value problem
has exactly one solution, which is given by y(t) = e
tA
y(0). Moreover, the
solution exists for all time t.
Proof. See Sections 1.31.4 of Perko [8]. The proof is not difﬁcult, but is best done
by introducing the concept of a matrix norm (which we shall not work with
here).
In the absence of a vector of initial conditions, the general solution of a constant
coefﬁcient system y
/
= Ay is y(t) = e
tA
u, where u is a vector of arbitrary
constants.
Example 2.1.11. Solve the initial value problem
dx
dt
= x −y x(0) = C
1
dy
dt
= 2x +4y y(0) = C
2
.
To do so, we write the system in matrix form
_
x
/
y
/
_
=
_
1 −1
2 4
_ _
x
y
_
.
The coefﬁcient matrix is the same matrix A used in the previous example, so we
have already computed its matrix exponential. According to Theorem 2.1.10, the
li near, constant coeffi ci ent systems 21
solution is
_
x(t)
y(t)
_
= e
tA
_
x(0)
y(0)
_
=
_
2e
2t
−e
3t
e
2t
−e
3t
−2e
2t
+2e
3t
−e
2t
+2e
3t
_ _
C
1
C
2
_
=
_
(2C
1
+ C
2
)e
2t
−(C
1
+ C
2
)e
3t
(−2C
1
−C
2
)e
2t
+ (2C
1
+2C
2
)e
3t
_
.
This represents the general solution of the system of odes. For the speciﬁc choice
of initial conditions x(0) = 8 and y(0) = 1, the solution would be
_
x(t)
y(t)
_
=
_
17e
2t
−9e
3t
−17e
2t
+18e
3t
_
.
2.1.2 Algebraic and Geometric Multiplicities of Eigenvalues. The process of
computing matrix exponentials can be tedious, especially for matrices which are
not diagonalizable over R. In preparation for things to come, we introduce some
notions from linear algebra which may not be familiar to all readers.
Deﬁnition 2.1.12. (Algebraic multiplicity.) Suppose λ
∗
is an eigenvalue of A and,
more speciﬁcally, that the factor (λ −λ
∗
) is repeated m times in the factorization
of the characteristic polynomial. We say that λ
∗
is an eigenvalue of algebraic
multiplicity m.
Example 2.1.13. If we ﬁnd that A has characteristic equation λ(λ−3)
4
(λ
2
+1) =
0, then λ = 0 has algebraic multiplicity 1 and λ = 3 has algebraic multiplicity
4. There are actually complex eigenvalues λ = ±i, both of which have algebraic
multiplicity 1. We will deal with complex eigenvalues later.
Deﬁnition 2.1.14. (Eigenspace and geometric multiplicity.) Suppose λ is an
eigenvalue of A. The eigenspace of λ is the span of all eigenvectors associated
with λ. The dimension of the eigenspace of λ is called the geometric multiplicity
of λ.
Example 2.1.15. Consider the matrices
A =
_
¸
_
1 1 0
0 1 0
0 0 2
_
¸
_
B =
_
¸
_
1 0 1
0 1 0
0 0 2
_
¸
_
.
22 homogeneous systems
You can check that both A and B have the same characteristic polynomial, namely
p(λ) = (λ −1)
2
(λ −2). Therefore, λ = 1 is an eigenvalue with algebraic multi
plicity 2, and λ = 2 is an eigenvalue with algebraic multiplicity 1. To calculate
the geometric multiplicities, we start with the matrix A and the eigenvalue λ = 1.
In this case,
A −λI =
_
¸
_
0 1 0
0 0 0
0 0 1
_
¸
_
,
which is row equivalent to
_
¸
_
0 1 0
0 0 1
0 0 0
_
¸
_
.
To ﬁnd eigenvectors, observe that solutions of (A −λI)v = 0 must satisfy
_
¸
_
0 1 0
0 0 1
0 0 0
_
¸
_
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
0
0
0
_
¸
_
.
Expanding reveals that v
2
= v
3
= 0, while v
1
is a free variable. The solutions of
the above system are
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
v
1
0
0
_
¸
_
= v
1
_
¸
_
1
0
0
_
¸
_
,
from which we conclude that
_
¸
_
1
0
0
_
¸
_
is an eigenvector corresponding to λ = 1. The span of this eigenvector is a one
dimensional subspace of R
3
, which means the eigenspace is onedimensional.
Equivalently, λ = 1 has geometric multiplicity 1. A similar calculation shows
that, for the matrix A, the geometric multiplicity of the eigenvalue λ = 2 is also
1.
li near, constant coeffi ci ent systems 23
Now for the matrix B, the eigenvalue λ = 1 gives rise to the matrix
B −λI =
_
¸
_
0 0 1
0 0 0
0 0 1
_
¸
_
.
This time, row reduction leads to a matrix which has two rows consisting entirely
of zeros. Eigenvectors must satisfy
_
¸
_
0 0 1
0 0 0
0 0 0
_
¸
_
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
0
0
0
_
¸
_
,
which implies that v
3
= 0 but that both v
1
and v
2
are free variables. Solutions of
(B −λI)v = 0 are of the form
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
v
1
v
2
0
_
¸
_
= v
1
_
¸
_
1
0
0
_
¸
_
+ v
2
_
¸
_
0
1
0
_
¸
_
.
Therefore, we have obtained a set of two linearly independent eigenvectors
_
¸
_
1
0
0
_
¸
_
and
_
¸
_
0
1
0
_
¸
_
,
which means that the eigenspace for λ = 1 is twodimensional. In other words,
for the matrix B, the geometric multiplicity of λ = 1 is 2. The reader can verify
that the geometric multiplicity of λ = 2 is 1.
The above example gives rise to several important observations, namely
The sum of the algebraic multiplicities of the eigenvalues always equals n, the
dimension of the underlying space. (In the above example, n = 3.)
In the example, the sum of the geometric multiplicities of the eigenvalues of A
is 2, whereas for the matrix B, the sum of the geometric multiplicities is 3.
In order to calculate the geometric multiplicities, it was actually not necessary
to compute the eigenvectors. The geometric multiplicity of an eigenvalue λ of
a matrix A is the same as the number of zero rows in the reduced row echelon
24 homogeneous systems
form of the matrix (A − λI). Since det(A − λI) = 0, the matrix (A − λI) is
singular, implying that there will be at least one row of zeros in the reduced
rowechelon form. Thus, every eigenvalue has a geometric multiplicity at least 1.
The following Lemma is useful in proving Theorem 2.1.17, which gives another
criterion for determining whether a matrix M is diagonalizable.
Lemma 2.1.16. Eigenvectors corresponding to different eigenvalues are linearly
independent.
Proof. We prove this statement for a set of 2 eigenvectors; the reader can extend
the proof to the general case. Let v
1
and v
2
be eigenvectors of a matrix A
corresponding to different eigenvalues λ
1
and λ
2
. Suppose indirectly that these
two eigenvectors are linearly dependent. Then there exists a constant c such that
v
2
= cv
1
. Moreover, since eigenvectors are nonzero, it must be the case that c ,= 0.
Multiplying both sides of the equation by A, we have Av
2
= cAv
1
. Equivalently,
λ
2
v
2
= cλ
1
v
1
. Replacing v
2
with cv
1
, this implies that λ
2
cv
1
= λ
1
cv
1
. But since
v
1
,= 0 and c ,= 0, this would mean that λ
2
= λ
1
, contradicting our assumption
that these were different eigenvalues.
Theorem 2.1.17. Suppose M is an n n matrix with real eigenvalues. Then M
is diagonalizable if and only if the sum of the geometric multiplicities of the
eigenvalues is equal to n. (Note: If an eigenvalue has algebraic multiplicity larger
than 1, we count its geometric multiplicity only once when forming the sum.)
Notice that, according to this Theorem, the matrix A in the previous example
is not diagonalizable, whereas the matrix B is diagonalizable because we can
produce a basis for R
3
consisting entirely of eigenvectors. By contrast, the
matrix A fails to be diagonalizable because it is “deﬁcient” in the sense that its
eigenvectors can only span a 2dimensional subspace of R
3
. In general, the sum
of the geometric multiplicities of the distinct eigenvalues can never exceed n, but
this sum can certainly be less than n (c.f., the matrix A in the previous example).
One nice consequence of Lemma 2.1.16 and Theorem 2.1.17 is the following
Corollary 2.1.18. If an n n matrix M has n distinct real eigenvalues, then M is
diagonalizable.
Example 2.1.19 (Triangular Matrices). If a square matrix M has no nonzero
entries above its main diagonal, then M is called lower triangular, and if it has no
nonzero entries below its main diagonal, then M is called upper triangular. More
li near, constant coeffi ci ent systems 25
rigorously, M is lower triangular if M
ij
= 0 whenever i < j, and M is upper
triangular if M
ij
= 0 whenever i > j. Writing down the characteristic equation
for a triangular matrix is easy. For example, suppose that A is uppertriangular
of the form
A =
_
¸
¸
¸
¸
¸
_
a
11
a
12
· · · a
1n
0 a
22
· · · a
2n
.
.
.
.
.
.
.
.
.
.
.
.
0 0 · · · a
nn
_
¸
¸
¸
¸
¸
_
.
Then the characteristic equation det(A −λI) = 0 is simply
(λ −a
11
)(λ −a
22
) · · · (λ −a
nn
) = 0,
which means that the eigenvalues of a triangular matrix are the entries on the main
diagonal. In light of Corollary 2.1.18, we immediately see that the triangular
matrix
_
¸
_
1 0 0
6 8 0
−6 3 −4
_
¸
_
is diagonalizable because its eigenvalues 1, 8, and −4 are real and distinct.
We now work through an example which connects much of the material in
this subsection.
Example 2.1.20. Solve the system
dx
1
dt
= x
1
+ x
3
x
1
(0) = 1
dx
2
dt
= x
2
+2x
3
x
2
(0) = 1
dx
3
dt
= 3x
3
x
3
(0) = 1.
Solution: In matrix notation, this system takes the form
_
¸
_
x
/
1
x
/
2
x
/
3
_
¸
_
=
_
¸
_
1 0 1
0 1 2
0 0 3
_
¸
_
_
¸
_
x
1
x
2
x
3
_
¸
_
.
26 homogeneous systems
Letting x denote the vector of unknowns and A the coefﬁcient matrix, the
solution of the system is x(t) = e
tA
x(0). To calculate the matrix exponential,
ﬁrst observe that since A is triangular, we may immediately conclude that the
eigenvalues are the entries on the main diagonal. The eigenvalue λ = 1 has
algebraic multiplicity 2. To ﬁnd eigenvectors associated with this eigenvalue, we
rowreduce the matrix (A −λI) = A − I:
A − I =
_
¸
_
0 0 1
0 0 2
0 0 2
_
¸
_
→
_
¸
_
0 0 1
0 0 0
0 0 0
_
¸
_
.
We immediately see that λ = 1 has geometric multiplicity 2, and we expect to
ﬁnd a set of two linearly independent eigenvectors for this eigenvalue. Solving
the system (A −λI)v = 0 as usual, we ﬁnd that v
3
= 0 while v
1
and v
2
are free.
Thus,
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
v
1
v
2
0
_
¸
_
= v
1
_
¸
_
1
0
0
_
¸
_
+ v
2
_
¸
_
0
1
0
_
¸
_
is the set of solutions of (A − I)v = 0, and we have obtained eigenvectors
_
¸
_
1
0
0
_
¸
_
and
_
¸
_
0
1
0
_
¸
_
.
The reader can show that the eigenvalue λ = 3 has geometric multiplicity 1 and
gives rise to an eigenvector
_
¸
_
1
2
1
1
_
¸
_
.
By Lemma 2.1.16, we know that the three eigenvectors we have produced are
linearly independent and form a basis for our space R
3
. This means that A is
diagonalizable, so form the matrix P by arranging the eigenvectors as columns:
P =
_
¸
_
1 0
1
2
0 1 1
0 0 1
_
¸
_
.
li near, constant coeffi ci ent systems 27
Then A = PDP
−1
, where D is the diagonal matrix
D =
_
¸
_
1 0 0
0 1 0
0 0 3
_
¸
_
and
P
−1
=
_
¸
_
1 0 −
1
2
0 1 −1
0 0 1
_
¸
_
.
Notice that D has the eigenvalues on its main diagonal, arranged in the same
order that the corresponding eigenvectors were arranged in P. The matrix
exponential e
tA
is given by
e
tA
= Pe
tD
P
−1
=
_
¸
_
1 0
1
2
0 1 1
0 0 1
_
¸
_
_
¸
_
e
t
0 0
0 e
t
0
0 0 e
3t
_
¸
_
_
¸
_
1 0 −
1
2
0 1 −1
0 0 1
_
¸
_
which, after some tedious matrix multiplication, simpliﬁes to
e
tA
=
_
¸
_
e
t
0 −
1
2
e
t
+
1
2
e
3t
0 e
t
−e
t
+e
3t
0 0 e
3t
_
¸
_
.
Finally, we multiply e
tA
by our vector of initial conditions to obtain the solution
of the initial value problem:
_
¸
_
x
1
(t)
x
2
(t)
x
3
(t)
_
¸
_
=
_
¸
_
e
t
0 −
1
2
e
t
+
1
2
e
3t
0 e
t
−e
t
+e
3t
0 0 e
3t
_
¸
_
_
¸
_
1
1
1
_
¸
_
=
_
¸
_
1
2
e
t
+
1
2
e
3t
e
3t
e
3t
_
¸
_
.
Observation. In the above example, notice that the exponential functions in
volved in the solution contain the eigenvalues in their exponents. Since λ = 1
and λ = 3 are the eigenvalues, the functions e
t
and e
3t
appear in the solutions.
Observation. The previous Observation suggests that, if eigenvalues are positive,
then the exponential functions in the solutions would increase without bound as
t → ∞. Negative eigenvalues would give rise to solutions which decay to 0 as
28 homogeneous systems
t →∞. For example, if λ = −4 is an eigenvalue, we would expect the decaying
exponential function e
−4t
to appear within the general solution.
Now that we know to expect exponential functions to serve as the “building
blocks” of our solutions, there is a much faster way of solving certain systems
y
/
= Ay without exponentiating the matrix A.
Proposition 2.1.21. Suppose that y
/
= Ay is a system of odes, where A is
an n n matrix which is diagonalizable (i.e., we can ﬁnd a set of n linearly
independent eigenvectors for A). Let λ
1
, λ
2
, . . . λ
n
denote the (possibly repeated)
eigenvalues, and let v
1
, v
2
, . . . v
n
denote their corresponding eigenvectors. Then
the general solution of the system y
/
= Ay is given by
y(t) = c
1
e
λ
1
t
v
1
+ c
2
e
λ
2
t
v
2
+ · · · + c
n
e
λ
n
t
v
n
. (2.9)
Proof. Suppose that y(t) is as in Equation (2.9). We must show that y(t) satisﬁes
y
/
= Ay. Differentiating with respect to t, we calculate
y
/
(t) = c
1
λ
1
e
λ
1
t
v
1
+ c
2
λ
2
e
λ
2
t
v
2
+ · · · + c
n
λ
n
e
λ
n
t
v
n
= c
1
e
λ
1
t
(λ
1
v
1
) + c
2
e
λ
2
t
(λ
2
v
2
) + · · · + c
n
e
λ
n
t
(λ
n
v
n
)
= c
1
e
λ
1
t
Av
1
+ c
2
e
λ
2
t
Av
2
+ · · · + c
n
e
λ
n
t
Av
n
= A
_
c
1
e
λ
1
t
v
1
+ c
2
e
λ
2
t
v
2
+ · · · + c
n
e
λ
n
t
v
n
_
= Ay(t).
Example 2.1.22. Solve the initial value problem
x
/
1
= −3x
1
+5x
2
x
1
(0) = 2
x
/
2
= x
1
+ x
2
x
2
(0) = −1.
Solution: In matrix notation, the system has the form x
/
= Ax, where the
coefﬁcient matrix A is given by
A =
_
−3 5
1 1
_
.
The characteristic polynomial is
det(A −λI) = (−3 −λ)(1 −λ) −5 = λ
2
+2λ −8 = (λ +4)(λ −2),
li near, constant coeffi ci ent systems 29
which means that the eigenvalues are λ = −4 and λ = 2. Because we have
distinct, real eigenvalues, Corollary 2.1.18 tells us that A is diagonalizable. The
reader can check that
_
1
1
_
and
_
−5
1
_
are eigenvectors corresponding to λ = 2 and λ = −4, respectively. Proposi
tion 2.1.21 tells us that the general solution of our system is given by
_
x
1
(t)
x
2
(t)
_
= c
1
e
2t
_
1
1
_
+ c
2
e
−4t
_
−5
1
_
,
where c
1
and c
2
are arbitrary constants. We determine the values of these
constants by using the initial conditions. Setting t = 0, the formula for the
general solution becomes
_
2
−1
_
= c
1
_
1
1
_
+ c
2
_
−5
1
_
=
_
1 −5
1 1
_ _
c
1
c
2
_
.
This is an inhomogeneous system of two equations in two unknowns, and basic
linear algebra techniques will lead you to the solution c
1
= −1/2 and c
2
= −1/2.
Finally, we see that the solution of the initial value problem is
_
x
1
(t)
x
2
(t)
_
=
_
−
1
2
e
2t
+
5
2
e
−4t
−
1
2
e
2t
−
1
2
e
−4t
_
.
2.1.3 Complex Eigenvalues. Unfortunately, not all matrices are diagonalizable,
and not all matrices have real eigenvalues. We will now learn how to exponentiate
matrices with complex conjugate eigenvalues. Recall that the imaginary unit i is
deﬁned according to the rule i
2
= −1. All complex numbers z can be written
in the form z = α + βi, where α and β are real numbers. Here, α is called the
real part of z, and β is called the imaginary part of z. Finally, we remark that the
complex numbers α + βi and α − βi are called complex conjugates of each other.
Example 2.1.23. The matrix
A =
_
1 −2
2 1
_
30 homogeneous systems
has characteristic polynomial det(A−λI) = (1 −λ)(1 −λ) +4 = λ
2
−2λ +5.
The roots of this polynomial are λ = 1 ± 2i, and hence our eigenvalues are
complex conjugates.
We still wish to solve constantcoefﬁcient systems of odes for which the
coefﬁcient matrix has complex conjugate eigenvalues. How can we exponentiate
such matrices? As a ﬁrst step towards answering this question, we tackle a
special case. Suppose
M =
_
0 −β
β 0
_
, (2.10)
where β is a nonzero real number. Notice that the characteristic equation for
M is given by λ
2
+ β
2
= 0. The roots of this equation are λ = ±βi, which are
pure imaginary numbers (i.e., the real part is zero).
Lemma 2.1.24. If M is the matrix in Equation (2.10), then
e
tM
=
_
cos βt −sin βt
sin βt cos βt
_
. (2.11)
Proof. The series representation for the matrix exponential
e
tM
=
∞
∑
k=0
t
k
k!
M
k
requires us to calculate powers of M. Fortunately, since
M
2
=
_
−β
2
0
0 −β
2
_
= −β
2
I,
computing higher powers of M is very straightforward. We focus our attention
on the (1, 1) entries of the matrices in the above summation. The upperleft entry
in the matrix e
tM
would be
t
0
0!
−
β
2
t
2
2!
+
β
4
t
4
4!
· · · =
∞
∑
k=0
(−1)
k
β
2k
t
2k
(2k)!
.
This is precisely the Maclaurin series for the function cos βt. A similar calculation
works for the other entries in the matrix.
li near, constant coeffi ci ent systems 31
Of course, we shall wish to exponentiate matrices with complex conjugate
eigenvalues which are not pure imaginary. In order to extend Lemma 2.1.24 to
handle more general matrices, we will need another Lemma:
Lemma 2.1.25. Suppose A and B are n n matrices. If the matrices commute
(AB = BA) then e
A+B
= e
A
e
B
.
Proof. See Section 1.3 of Perko [8].
Warning: If matrices A and B do not commute, then we cannot conclude that
e
A+B
= e
A
e
B
. The reader is encouraged to test this, using the matrices
A =
_
1 0
0 0
_
and B =
_
0 1
0 0
_
.
Lemmas 2.1.24 and 2.1.25 can be combined to prove
Proposition 2.1.26. Suppose
A =
_
α −β
β α
_
,
where α and β are real numbers. Then
e
tA
=
_
e
αt
cos βt −e
αt
sin βt
e
αt
sin βt e
αt
cos βt
_
= e
αt
_
cos βt −sin βt
sin βt cos βt
_
.
Proof. We write A as a sum of two matrices:
A =
_
α 0
0 α
_
+
_
0 −β
β 0
_
= B + M,
where M is the matrix from Lemma 2.1.24. Clearly B and M commute because
B = αI is simply a scalar multiple of the identity matrix. Therefore, Lemma 2.1.25
tells us that
e
tA
= e
t(B+M)
= e
tB+tM
= e
tB
e
tM
.
We know how to compute both of these matrix exponentials:
e
tB
=
_
e
αt
0
0 e
αt
_
= e
αt
I and e
tM
=
_
cos βt −sin βt
sin βt cos βt
_
.
32 homogeneous systems
Therefore,
e
tA
= e
tB
e
tM
= e
αt
_
cos βt −sin βt
sin βt cos βt
_
,
as claimed.
Remark. In the above Proposition, you may check that the characteristic equation
for A is given by
λ
2
−2αλ + (α
2
+ β
2
) = 0,
and has roots λ = α ± βi.
We now show that every 2 2 matrix with complex conjugate eigenvalues
α ± βi is similar to a matrix of the form in Proposition 2.1.26. We will exploit this
fact to learn how to exponentiate any matrix with complex eigenvalues.
Deﬁnition 2.1.27. If A is a matrix with complex conjugate eigenvalues α ± βi,
then the matrix
M =
_
α −β
β α
_
is called the real canonical form for A.
The next Theorem states that every 2 2 matrix with complex conjugate eigen
values is similar to its real canonical form. Since we know how to exponentiate
matrices in real canonical form, we should be able to solve systems of constant
coefﬁcient odes in cases where the coefﬁcient matrix has complex conjugate
eigenvalues. Notice that we do not attempt to diagonalize such matrices, be
cause we would prefer to work with matrices containing real entries (to facilitate
ﬁnding realvalued solutions of odes).
Theorem 2.1.28. Suppose A is a 2 2 matrix with eigenvalues α ± βi, and
suppose w is a complex eigenvector corresponding to the eigenvalue α + βi.
Write the eigenvector as w = u +iv, where u and v are real vectors, and form
the matrix P = [vu]. Then
A = P
_
α −β
β α
_
P
−1
.
In other words, A is similar to a real canonical form.
li near, constant coeffi ci ent systems 33
Theorem 2.1.28 tells us exactly how to put a matrix in real canonical form, as
we illustrate in the following example.
Example 2.1.29. Solve the system
dx
1
dt
= x
1
−12x
2
dx
2
dt
= 3x
1
+ x
2
.
Solution: As usual, we write the system in matrix form
_
x
/
1
x
/
2
_
=
_
1 −12
3 1
_ _
x
1
x
2
_
.
Letting A denote the coefﬁcient matrix, the characteristic equation is λ
2
−2λ +
37 = 0, which has roots λ = 1 ± 6i. According to Theorem 2.1.28, we should
ﬁnd a complex eigenvector for the eigenvalue with positive imaginary part:
λ = 1 +6i. We calculate
A −λI =
_
1 −(1 +6i) −12
3 1 −(1 +6i)
_
=
_
−6i −12
3 −6i
_
.
When computing the reduced rowechelon form of this 2 2 matrix, we may
immediately replace one of the rows with zeros—we know the matrix is singular
because λ was an eigenvalue. This saves us a bit of time with elementary row
operations, because it is not necessary that we notice that Row 1 is −2i times
Row 2. The reduced rowechelon form is easy to calculate:
_
−6i −12
3 −6i
_
→
_
0 0
3 −6i
_
→
_
3 −6i
0 0
_
→
_
1 −2i
0 0
_
.
We seek a complex eigenvector w which satisﬁes (A −λI)w = 0, so set
_
1 −2i
0 0
_ _
w
1
w
2
_
=
_
0
0
_
.
We ﬁnd that w
1
−2iw
2
= 0, so w
1
= 2iw
2
and we treat w
2
as a free variable.
Thus,
_
w
1
w
2
_
=
_
2iw
2
w
2
_
= w
2
_
2i
1
_
.
34 homogeneous systems
Next, the theorem says we should write this (complex) eigenvector in terms of
its real and imaginary parts:
_
2i
1
_
= w = u +iv =
_
0
1
_
+ i
_
2
0
_
.
Form the matrix P = [vu] by placing the imaginary part of the eigenvector w in
the ﬁrst column, and the real part of w in the second column:
P =
_
2 0
0 1
_
.
This matrix is particularly easy to invert:
P
−1
=
_
1
2
0
0 1
_
.
Since our eigenvalue was λ = 1 + 6i, the notation used in the statement of
Theorem 2.1.28 says that α = 1 and β = 6. The reader can verify that A = PMP
−1
,
where
M =
_
1 −6
6 1
_
is in real canonical form. It follows that e
tA
= Pe
tM
P
−1
, and by Proposition 2.1.26,
we have
e
tA
=
_
2 0
0 1
_
e
t
_
cos 6t −sin6t
sin6t cos 6t
_ _
1
2
0
0 1
_
= e
t
_
cos 6t −2 sin6t
1
2
sin6t cos 6t
_
.
Finally, since no initial conditions were speciﬁed, we obtain the general solution
by multiplying e
tA
by a vector of arbitrary constants:
_
x
1
(t)
x
2
(t)
_
= e
t
_
cos 6t −2 sin6t
1
2
sin6t cos 6t
_ _
c
1
c
2
_
=
_
c
1
e
t
cos 6t −2c
2
e
t
sin6t
1
2
c
1
e
t
sin6t + c
2
e
t
cos 6t
_
.
Observation. Complex conjugate eigenvalues give rise to oscillatory solutions.
li near, constant coeffi ci ent systems 35
Example 2.1.30. Solve the system x
/
= Ax where
A =
_
¸
_
3 −2 0
1 1 0
0 0 4
_
¸
_
and x(0) =
_
¸
_
1
2
3
_
¸
_
.
Cool observation: The matrix A has a block diagonal structure
A =
_
¸
¸
_
3 −2 0
1 1 0
0 0 4
_
¸
¸
_
.
The upperleft corner is a 2 2 block, the bottomright corner is a 1 1 block, and
all other entries are zero. The reader is encouraged to prove that the exponential
of a blockdiagonal matrix is obtained by exponentiating each diagonal block
separately. In the present case, the only part that will require any work is the
2 2 block. To ﬁnd the eigenvalues of A, we calculate
det(A −λI) =
¸
¸
¸
¸
¸
¸
¸
3 −λ −2 0
1 1 −λ 0
0 0 4 −λ
¸
¸
¸
¸
¸
¸
¸
.
Performing a cofactor expansion using the last row, this determinant simpliﬁes
to
det(A −λI) = (4 −λ)
¸
¸
¸
¸
¸
3 −λ −2
1 1 −λ
¸
¸
¸
¸
¸
= (4 −λ)(λ
2
−4λ +5).
Thus, λ = 4 is one eigenvalue, and the other two are λ = 2 ± i. You can show
that
_
¸
_
0
0
1
_
¸
_
is an eigenvector corresponding to λ = 4. As in the previous example, we
must ﬁnd a complex eigenvector for the eigenvalue with positive imaginary part
(λ = 2 +i):
A −λI =
_
¸
_
1 −i −2 0
1 −1 −i 0
0 0 2 −i
_
¸
_
.
36 homogeneous systems
Clearly the third row is independent of the others. However, since λ = 2 +i is
an eigenvalue, we know that these rows must be linearly dependent. This means
that rows 1 and 2 must form a linearly dependent set and, since there are only
two rows, must be scalar multiples of each other. Row reduction will annihilate
one of these two rows—we will replace the ﬁrst row with zeros since the second
row looks more convenient to work with. The reduced rowechelon form for
(A −λI) is
_
¸
_
1 −1 −i 0
0 0 1
0 0 0
_
¸
_
.
Eigenvectors w must satisfy
_
¸
_
1 −1 −i 0
0 0 1
0 0 0
_
¸
_
_
¸
_
w
1
w
2
w
3
_
¸
_
=
_
¸
_
0
0
0
_
¸
_
,
from which we conclude that w
1
+ (−1 −i)w
2
= 0 and w
3
= 0. Treating w
2
as
the free variable, we ﬁnd that w
1
= (1 +i)w
2
. To ﬁnd our complex eigenvector,
we write
_
¸
_
w
1
w
2
w
3
_
¸
_
=
_
¸
_
(1 +i)w
2
w
2
0
_
¸
_
= w
2
_
¸
_
1 +i
1
0
_
¸
_
.
As before, we follow the instructions of Theorem 2.1.28 by splitting w into its
real and imaginary parts:
_
¸
_
1 +i
1
0
_
¸
_
= w = u +iv =
_
¸
_
1
1
0
_
¸
_
+ i
_
¸
_
1
0
0
_
¸
_
.
Next, we form the matrix P by arranging the vectors as suggested by Theo
rem 2.1.28. Column 1 should be the vector v, Column 2 should be u, and
Column 3 should be the eigenvector corresponding to the real eigenvalue λ = 4.
In summary,
P =
_
¸
_
1 1 0
0 1 0
0 0 1
_
¸
_
.
li near, constant coeffi ci ent systems 37
It is easily checked that
P
−1
=
_
¸
_
1 −1 0
0 1 0
0 0 1
_
¸
_
and that A = PMP
−1
where
M =
_
¸
_
2 −1 0
1 2 0
0 0 4
_
¸
_
.
Notice that M has the same blockdiagonal form as the original matrix A. The
upperleft block of M contains the real canonical form of the upperleft block of
A.
Our last step is to calculate e
tA
and to multiply this matrix by the vector of
initial conditions. Using Proposition 2.1.26, we compute
e
tA
= Pe
tM
P
−1
= P
_
¸
_
e
2t
cos t −e
2t
sin t 0
e
2t
sin t e
2t
cos t 0
0 0 e
4t
_
¸
_
P
−1
,
where we have exponentiated the two diagonal blocks separately. After evaluat
ing the above matrix product (which is tedious), multiply the result by the vector
of initial conditions to obtain
_
¸
_
x
1
(t)
x
2
(t)
x
3
(t)
_
¸
_
= e
tA
_
¸
_
1
2
3
_
¸
_
=
_
¸
_
e
2t
(cos t −3 sin t)
e
2t
(2 cos t −sin t)
3e
4t
_
¸
_
.
2.1.4 Repeated Eigenvalues and NonDiagonalizable Matrices. We now turn
our attention to devising a method for exponentiating matrices that are not
diagonalizable. In order for a matrix to be diagonalizable, we need the geometric
multiplicities of the different eigenvalues to sum to n, the dimension of the
underlying space. Sometimes, eigenvalues with high algebraic multiplicity may
have low geometric multiplicity.
38 homogeneous systems
Example 2.1.31. Compute e
tA
, where
A =
_
¸
_
0 1 0
0 0 1
0 0 0
_
¸
_
.
Since A is uppertriangular, we immediately see that λ = 0 is an eigenvalue
of algebraic multiplicity 3. To ﬁnd eigenvectors, we must solve (A −λI)v = 0
which, in this case, simpliﬁes to Av = 0. The matrix A is already in reduced
rowechelon form, and the vectors v which satisfy Av = 0 are of the form
v
2
= v
3
= 0, with v
1
free. Thus,
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
v
1
0
0
_
¸
_
= v
1
_
¸
_
1
0
0
_
¸
_
,
and we see that the eigenspace is only onedimensional. Since our only eigen
value λ = 0 has geometric multiplicity 1, Theorem 2.1.17 tells us that A is not
diagonalizable.
In order to compute the matrix exponential, we resort to a direct use of the
deﬁnition
e
tA
=
∞
∑
k=0
t
k
k!
A
k
= I + tA +
t
2
2!
A
2
+ · · ·
Fortunately, for this particular matrix, the powers are not so difﬁcult to calculate.
You can show that
A
2
=
_
¸
_
0 0 1
0 0 0
0 0 0
_
¸
_
and that A
3
is the zero matrix. It follows that the matrix A
k
= 0 if k ≥ 3, which
means that the matrix exponential is actually a ﬁnite sum
e
tA
= I + tA +
t
2
2!
A
2
=
_
¸
_
1 t
1
2
t
2
0 1 t
0 0 1
_
¸
_
.
This example motivates the following deﬁnition.
li near, constant coeffi ci ent systems 39
Deﬁnition 2.1.32. Let k be a positive integer. An n n matrix N is called nilpotent
of order k if N
k−1
,= 0 but N
k
= 0.
In the above example, A is nilpotent of order 3. The fact that a nonzero matrix
can be nilpotent is quite different from what the reader may be accustomed to.
Certainly the concept of nilpotency does not apply to real numbers—if α is a real
number and k ≥ 1, it is impossible for α
k
= 0 unless α = 0.
The matrix exponential of a nilpotent matrix is represented by a ﬁnite sum, a
fact that we alluded to in the previous example and now state formally.
Lemma 2.1.33. If N is nilpotent of order k, then
e
tN
= I + tN +
t
2
2!
N
2
+ · · · +
t
k−1
(k −1)!
N
k−1
=
k−1
∑
j=0
t
j
j!
N
j
.
Proof. Since N is nilpotent of order k, we know that N
m
= 0 whenever m ≥ k.
This means that we can “drop” every term of index k or higher in the series
representation of e
tN
.
Before we present a method for exponentiating nondiagonalizable matrices,
we need one last deﬁnition from linear algebra.
Deﬁnition 2.1.34. Suppose A is an n n matrix and that λ is an eigenvalue of
algebraic multiplicity m ≤ n. Then for each k = 1, 2, . . . m, any nonzero solution
v of the equation (A −λI)
k
v = 0 is called a generalized eigenvector of A.
The following Theorem tells us how to decompose a nondiagonalizable matrix
as a sum of two types of matrices that we know how to exponentiate.
Theorem 2.1.35. (Simple Jordan Decompositions): Suppose that A is an n
n matrix with real eigenvalues λ
1
, λ
2
, . . . λ
n
, repeated according to algebraic
multiplicity. Then there exists a family of n generalized eigenvectors v
1
, v
2
, . . . v
n
such that
The matrix P = [v
1
v
2
 · · · v
n
] is invertible.
A = S + N where N is nilpotent and S is diagonalizable. Speciﬁcally, if we
deﬁne D = diag{λ
1
, λ
2
, . . . λ
n
}, then S = PDP
−1
.
S and N commute: SN = NS.
We illustrate the usefulness of this Theorem via an example.
40 homogeneous systems
Example 2.1.36. Compute e
tA
where
A =
_
¸
_
1 0 0
−1 2 0
1 1 2
_
¸
_
.
Solution: Since A is lowertriangular, the eigenvalues are the entries on the
main diagonal. You can check that
_
¸
_
1
1
−2
_
¸
_
is an eigenvector for the eigenvalue λ = 1. The other eigenvalue λ = 2 has
algebraic multiplicity 2. To compute corresponding eigenvectors, we rowreduce
(A −λI) as
A −2I =
_
¸
_
−1 0 0
−1 0 0
1 1 0
_
¸
_
→
_
¸
_
1 0 0
0 1 0
0 0 0
_
¸
_
.
Hence, solutions of (A −2I)v = 0 have the form v
1
= v
2
= 0, with v
3
free. It
follows that
_
¸
_
0
0
1
_
¸
_
is an eigenvector for λ = 2. Unfortunately, since the geometric multiplicity of
this eigenvalue is only 1, we have failed to produce a set of 3 linearly indepen
dent eigenvectors for the matrix A, which means that A is not diagonalizable.
Theorem 2.1.35 suggests that we compute a generalized eigenvector for λ = 2.
We must solve (A −2I)
2
v = 0, or equivalently
_
¸
_
1 0 0
1 0 0
−2 0 0
_
¸
_
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
0
0
0
_
¸
_
.
li near, constant coeffi ci ent systems 41
Rowreduction gives
_
¸
_
1 0 0
0 0 0
0 0 0
_
¸
_
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
0
0
0
_
¸
_
,
which has solution v
1
= 0 with both v
2
and v
3
free. Thus, solutions of (A −
2I)v = 0 are
_
¸
_
v
1
v
2
v
3
_
¸
_
=
_
¸
_
0
v
2
v
3
_
¸
_
= v
2
_
¸
_
0
1
0
_
¸
_
+ v
3
_
¸
_
0
0
1
_
¸
_
.
We recognize
_
¸
_
0
0
1
_
¸
_
as the eigenvector that we had already produced, which means that we can take
_
¸
_
0
1
0
_
¸
_
as our generalized eigenvector. Next, we form the matrix
P =
_
¸
_
1 0 0
1 0 1
−2 1 0
_
¸
_
.
Notice that Column 1 contains the eigenvector for the eigenvalue λ = 1, Column 2
contains the eigenvector for λ = 2, and Column 3 contains the generalized
eigenvector for λ = 2. Inverting P, we ﬁnd that
P
−1
=
_
¸
_
1 0 0
2 0 1
−1 1 0
_
¸
_
.
42 homogeneous systems
Let D denote the diagonal matrix D = diag{1, 2, 2} consisting of the eigenvalues,
arranged in an order corresponding to the order we used for the columns of P.
The matrix S in the statement of Theorem 2.1.35 is given by
S = PDP
−1
=
_
¸
_
1 0 0
−1 2 0
2 0 2
_
¸
_
,
from which we deﬁne
N = A −S =
_
¸
_
0 0 0
0 0 0
−1 1 0
_
¸
_
.
The Theorem says that N should be a nilpotent matrix—indeed, the reader
can check that N
2
= 0. Importantly, the Theorem also guarantees that S and N
commute, which means that e
tA
= e
t(S+N)
= e
tS
e
tN
. Since S is diagonalizable and
N is nilpotent, we know how to exponentiate both of these matrices. According
to Lemma 2.1.33, since N is nilpotent of order 2, we have
e
tN
= I + tN =
_
¸
_
1 0 0
0 1 0
−t t 1
_
¸
_
.
Moreover,
e
tS
= Pe
tD
P
−1
= P
_
¸
_
e
t
0 0
0 e
2t
0
0 0 e
2t
_
¸
_
P
−1
.
Finally, since e
tA
= e
tS
e
tN
, after tedious matrix multiplication we ﬁnd that
e
tA
=
_
¸
_
e
t
0 0
e
t
−e
2t
e
2t
0
−2e
t
+ (2 −t)e
2t
te
2t
e
2t
_
¸
_
.
Observation. Repeated real eigenvalues can cause the matrix exponential to
contain products of polynomials with exponentials; e.g., (2 −t)e
2t
or t
7
e
−5t
.
li near, constant coeffi ci ent systems 43
Example 2.1.37. Solve the initial value problem
x
/
1
= x
2
x
1
(0) = 2
x
/
2
= −9x
1
−6x
2
x
2
(0) = 1.
Solution: If we write the system in matrix form, the coefﬁcient matrix is
A =
_
0 1
−9 −6
_
.
The characteristic equation is λ
2
+ 6λ + 9 = 0, which factors as (λ + 3)
2
= 0.
Thus, λ = −3 is an eigenvalue of algebraic multiplicity 2. The matrix A −λI
row reduces to
_
3 1
0 0
_
,
which means that eigenvectors v must satisfy the equation
_
3 1
0 0
_ _
v
1
v
2
_
=
_
0
0
_
.
This implies that 3v
1
+ v
2
= 0 and, treating v
2
as the free variable, we may write
v
1
= −
1
3
v
2
. Solutions of (A −λI)v = 0 have the form
_
v
1
v
2
_
= v
2
_
−
1
3
1
_
,
which means that our eigenspace is only onedimensional. Theorem 2.1.35
suggests that we ﬁnd a generalized eigenvector by solving (A − λI)
2
v = 0.
Interestingly,
(A −λI)
2
= 0 ←− Remember this later!!!
(Used properly, this observation will allow for a much quicker solution to our
problem. We will see how after going through the tedious motions of Theo
rem 2.1.35.) Consequently, if we attempt to solve (A − λI)
2
v = 0, then both
v
1
and v
2
are free variables (i.e., any vector will satisfy this equation). For our
generalized eigenvector, we may select any vector which is independent of the
44 homogeneous systems
eigenvector we already computed. It is convenient to choose
_
0
1
_
as our generalized eigenvector, and form the matrix
P =
_
−
1
3
0
1 1
_
.
Again, notice that Column 1 contains our eigenvector and Column 2 contains
our generalized eigenvector. You can check that
P
−1
=
_
−3 0
3 1
_
.
Letting D = diag{−3, −3} denote the diagonal matrix with our repeated eigen
value as its entries, we compute the matrix S in the statement of Theorem 2.1.35:
S = PDP
−1
=
_
−
1
3
0
1 1
_ _
−3 0
0 −3
_ _
−3 0
3 1
_
=
_
−3 0
0 −3
_
.
(Again, we make the curious observation that S is not simply diagonalizable—it’s
diagonal!). Theorem 2.1.35 tells us that
N = A −S =
_
3 1
−9 −3
_
should be nilpotent, and you can check that N
2
= 0. We have A = S + N where
S and N commute, which means that e
tA
= e
tN
e
tS
. Since N is nilpotent of
order 2, Lemma 2.1.33 tells us that e
tN
= I + tN. Thus,
e
tA
= [I + tN]e
tS
=
_
1 +3t t
−9t 1 −3t
_ _
e
−3t
0
0 e
−3t
_
=
_
(1 +3t)e
−3t
te
−3t
−9te
−3t
(1 −3t)e
−3t
_
.
li near, constant coeffi ci ent systems 45
Finally, we multiply e
tA
by our vector of initial conditions to get the overall
solution
_
x
1
(t)
x
2
(t)
_
=
_
(1 +3t)e
−3t
te
−3t
−9te
−3t
(1 −3t)e
−3t
_ _
2
1
_
=
_
(2 +7t)e
−3t
(1 −21t)e
−3t
_
.
Remark. As soon as we noticed that (A −λI)
2
= 0 we could have saved lots of
time. The whole purpose of Theorem 2.1.35 is to decompose a matrix A as a
sum of a diagonalizable matrix S and a nilpotent matrix N. So upon observing
that A −λI is nilpotent (of order 2), we could immediately deﬁne A −λI = N.
According to the notation of the Theorem, this would mean S = λI. We have
written A = S + N where S is diagonal, N is nilpotent, and SN = NS. Hence,
we could have immediately computed the matrix exponential for A without any
further effort!
The techniques we have learned so far can be combined to handle cases
we have not considered (such as repeated complex conjugate eigenvalues). In
principle, it is possible to solve any linear constantcoefﬁcient system. We have
not attempted to state the most general possible theorems regarding solutions
of constantcoefﬁcient systems. Instead, we will move ahead to a topic that will
help us visualize the solutions.
2.2. Phase Portraits and Planar Systems
We now introduce the concept of the phase plane, a useful way of visualizing
solutions of planar systems (two dependent variables) of odes. We begin by
considering the system
_
x
/
1
x
/
2
_
=
_
0 −1
1 0
_ _
x
1
x
2
_
.
Observe that the coefﬁcient matrix A is already in real canonical form, and
e
tA
=
_
cos t −sin t
sin t cos t
_
.
If we were interested in the solution satisfying initial conditions x
1
(0) = 1 and
x
2
(0) = 0, we would multiply e
tA
by the vector [1 0] to obtain x
1
(t) = cos t
46 phase portrai ts and planar systems
x
1
(t)
x
2
(t)
1
0
−1
t
Figure 2.2. Graphs of x
1
(t) = cos t and x
2
(t) = sin t.
and x
2
(t) = sin t. The most natural way to visualize these solutions is to simply
graph both x
1
and x
2
as a function of t as in Figure 2.2. An alternate way to
visualize the solution is to plot x
2
(t) versus x
1
(t) as a curve parametrized by
time t. Such a plot is given in Figure 2.3, which illustrates the parametrized
curve (cos t, sin t) for t ∈ [0, 2π]. In this case, the parametrized curve is a circle
of radius 1 because x
1
(t) and x
2
(t) satisfy the relationship
x
2
1
+ x
2
2
= cos
2
(t) +sin
2
(t) = 1
for all t. Notice that as t increases, the curve is traversed in the counterclockwise
direction.
Other initial conditions would give rise to other parametric curves, giving a
visual representation of the “ﬂow” of the system. A parametric plot of x
2
(t)
versus x
1
(t) for various choices of initial conditions is called a phase portrait or
phase plane diagram for the system.
We now discuss three canonical examples of phase portraits for planar systems,
and then explain how these special cases can be used to sketch phase portraits
for more general planar systems. From the previous subsections, we know that
every 2 2 matrix A can be written in the form A = PMP
−1
, where M has one
of three forms:
li near, constant coeffi ci ent systems 47
x
2
x
1
Figure 2.3. Parametric plot of x
2
(t) = sin t versus x
1
(t) = cos t for
t ∈ [0, 2π].
If A is diagonalizable, then
M =
_
α 0
0 β
_
.
If A is nondiagonalizable and has a real, repeated eigenvalue α, then
M =
_
α 1
0 α
_
.
If A has complexconjugate eigenvalues α ± βi, then
M =
_
α −β
β α
_
.
Case 1: Suppose x
/
= Mx where M = diag{α, β}. The general solution is
x(t) = e
tM
x(0); i.e.,
_
x
1
(t)
x
2
(t)
_
=
_
e
αt
0
0 e
βt
_ _
c
1
c
2
_
=
_
c
1
e
αt
c
2
e
βt
_
.
The exponential solutions x
1
= c
1
e
αt
and x
2
= c
2
e
βt
either grow or decay
depending upon the signs of α and β. Suppose that both α and β are positive.
Then both e
αt
and e
βt
will increase without bound as t → ∞. We can sketch
48 phase portrai ts and planar systems
x
1
x
2
x
2
x
1
x
2
x
1
Figure 2.4. Phase portraits corresponding to Case 1 in which both α
and β are positive. Left panel: 0 < α < β. Middle panel: 0 < α = β.
Right panel: 0 < β < α.
the phase portrait by eliminating the parameter t. Notice that x
β
1
= c
β
1
e
αβt
and
x
α
2
= c
α
2
e
αβt
. Assuming that c
1
,= 0, taking the ratio gives
x
α
2
x
β
1
=
c
α
2
c
β
1
= c,
where c is a constant. By algebra, we have x
2
= cx
β/α
1
, and graphing such power
functions is easy. Figure 2.4 shows the possible phase portraits in cases where
both α and β are positive. Notice that all trajectories point outward from the
origin. In particular, if β = α (and thus β/α = 1), the trajectories in the phase
plane are straight lines directed outward from the origin.
If both α and β are negative, then both x
1
and x
2
will decay to zero exponen
tially fast as t → ∞. As indicated in Figure 2.5, the phase portraits are similar
to the previous ones, except that all trajectories are directed inward towards the
origin.
A more interesting situation occurs when α and β have different signs. In this
case, we have exponential growth for one of our variables and exponential decay
for the other variable. The possible phase portraits are shown in Figure 2.6.
The various phase portraits in Figures 2.4–2.6 all have one thing in common:
the special solution corresponding to c
1
= c
2
= 0 is a constant solution in which
we stay “stuck” at the origin in the phase plane for all time t. Constant solutions
of odes will be of particular interest to us, and we give them a special name.
Deﬁnition 2.2.1. An equilibrium of a system of odes is a solution that is constant.
That is, all dependent variables are constant for all time t.
li near, constant coeffi ci ent systems 49
x
1
x
2
x
2
x
1
x
2
x
1
Figure 2.5. Phase portraits corresponding to Case 1 in which both α
and β are negative. Left panel: β < α < 0. Middle panel: α = β < 0.
Right panel: α < β < 0.
x
1
x
2
x
1
x
2
Figure 2.6. Phase portraits corresponding to Case 1 in which both α and
β have different signs. Left panel: β < 0 < α. Right panel: α < 0 < β.
Observation. For linear, constantcoefﬁcient systems x
/
= Ax of odes, x = 0 is
always an equilibrium solution. In terms of the phase portrait, this means that
the origin always corresponds to an equilibrium solution.
One way of further classifying equilibria is provided by the phase portraits
discussed above. Namely,
Deﬁnition 2.2.2. In Figure 2.4, the origin is called an unstable node—all trajec
tories point outward from the origin. In Figure 2.5, the origin is called a stable
node—all trajectories point inward to the origin. In Figure 2.6, the origin is called
a saddle.
Later, we will actually restate this Deﬁnition in a more precise way. We remark
that saddle equilibria are different from stable/unstable nodes in that only certain,
special trajectories actually touch the equilibrium point. In Figure 2.6, these are
the trajectories which lie along the x
1
and x
2
axes. The four trajectories which
50 phase portrai ts and planar systems
approach the origin as t → ±∞ are called separatrices. (Again, we will make this
deﬁnition more precise later.)
Case 2: Suppose x
/
= Mx where M is the nondiagonalizable matrix
M =
_
α 1
0 α
_
with α as a repeated, real eigenvalue. The phase portrait is more difﬁcult to plot
without computer assistance. To solve the system of odes, we decompose M as
the sum of a diagonal matrix and a nilpotent matrix:
M = D + N =
_
α 0
0 α
_
+
_
0 1
0 0
_
.
Clearly D and N commute since D is a scalar multiple of the identity matrix,
and you can also verify that N is nilpotent of order 2. It follows that
e
tM
= e
tD
e
tN
= e
tD
[I + tN] =
_
e
αt
0
0 e
αt
_ _
1 t
0 1
_
=
_
e
αt
te
αt
0 e
αt
_
.
The general solution of the system is therefore x(t) = e
tM
x(0), or
_
x
1
(t)
x
2
(t)
_
=
_
e
αt
te
αt
0 e
αt
_ _
c
1
c
2
_
=
_
c
1
e
αt
+ c
2
te
αt
c
2
e
αt
_
.
These solutions give a parametrization for the solution curves in the phase
portrait, which appears in Figure 2.7. If α < 0, the trajectories approach the
origin as t → ∞, and if α > 0, the trajectories are oriented outward from the
origin. Notice that, unlike the case in which the origin is a saddle equilibrium,
there is only one separatrix in this case (the x
1
axis). In the left panel of Figure 2.7,
the origin is a stable node, and in the right panel the origin is an unstable node.
We remark that the characteristic equation for M is (λ −α)
2
= 0, which is
a “critical” case in the sense that we are just on the verge of having complex
conjugate roots. Indeed, quadratic equations with negative discriminants have
complex conjugate solutions, and repeated real roots occur when the discriminant
is 0.
li near, constant coeffi ci ent systems 51
x
1
x
2
x
2
x
1
Figure 2.7. Phase portraits corresponding to Case 2, in which M has a
repeated eigenvalue. Left panel: α = −1. Right panel: α = 1.
Case 3: Suppose x
/
= Mx where M is in real canonical form
M =
_
α −β
β α
_
.
We know that
e
tM
= e
αt
_
cos βt −sin βt
sin βt cos βt
_
,
which means that the general solution of the system is
_
x
1
(t)
x
2
(t)
_
= e
tM
_
c
1
c
2
_
=
_
c
1
e
αt
cos βt −c
2
e
αt
sin βt
c
1
e
αt
sin βt + c
2
e
αt
cos βt
_
.
Solutions involve combinations of exponential functions e
αt
with periodic func
tions sin βt and cos βt. If α < 0, the exponential factors will decay as t → ∞,
while the other factors simply oscillate. If we graph some solution curves in the
phase portrait, we will see that trajectories spiral inward to the origin. Likewise,
if α > 0, trajectories will spiral outward from the origin. Finally, if α = 0, the
exponential factor is constant and the solutions are purely oscillatory. In this
case, the phase portrait consists of concentric circular trajectories.
For matrices M in real canonical form
1
, the orientation of the trajectories can be
determined from the sign of β: clockwise if β < 0 and counterclockwise if β > 0.
These various possibilities are summarized in Figure 2.8. The equilibrium at the
1
If a matrix with complex conjugate eigenvalues is not in real canonical form, one simple way to
determine the orientation of the trajectories is to plot a few “slope ﬁeld vectors” as illustrated in the
next example.
52 phase portrai ts and planar systems
origin is called a stable focus if trajectories spiral inward (α < 0), an unstable focus
if trajectories spiral outward (α > 0), and a center if trajectories form concentric,
closed curves (α = 0).
Remark. Above, we did not consider the possibility that the matrix M in the
equation x
/
= Mx has λ = 0 as an eigenvalue. In such cases, the origin is called
a degenerate equilibrium. Notice that if λ = 0 is an eigenvalue, then det M = 0,
which means that M is a singular matrix. It follows that the solutions of the
equation Mx = 0 form a subspace of dimension at least 1, and any vector x in
this subspace would be an equilibrium for our system of odes. In the remainder
of the course, we will typically work only with systems which have isolated
equilibrium points (deﬁned later), as opposed to systems with inﬁnitely many
equilibrium points.
It is also worth noting that for the autonomous systems of odes that we have
considered, solution trajectories in the phase portrait cannot intersect each other. Can
you explain why?
Suppose a planar system of odes has a coefﬁcient matrix A which is not
in one of the three canonical forms we discussed above. To sketch the phase
portrait, we need to determine which canonical form A is similar to, and this
is accomplished by ﬁnding the eigenvalues and eigenvectors. If A is any 2 2
matrix with nonzero eigenvalues, then the associated phase portrait is always a
“skewed” version of one of the portraits in the above ﬁgures. We illustrate this
via two examples.
Example 2.2.3. Sketch the phase portrait for x
/
= Ax, where
A =
_
0 −4
1 0
_
.
Solution: The characteristic equation is λ
2
+ 4 = 0, from which we infer that
the eigenvalues are complex conjugate: λ = ±2i. We will put A in real canonical
form by ﬁnding an eigenvector for λ = 2i, the eigenvalue with positive imaginary
part. Rowreducing the matrix (A −λI) gives
A −λI =
_
−2i −4
1 −2i
_
−→
_
1 −2i
0 0
_
.
li near, constant coeffi ci ent systems 53
x
2
x
1
x
2
x
1
x
2
x
1
x
2
x
1
x
2
x
2
x
1
x
1
(a) (b)
(c)
(d)
(e) (f)
Figure 2.8. Phase portraits corresponding to Case 3, in which M has a
complex conjugate eigenvalues.
(a) Origin is a stable focus (α < 0) with trajectories oriented counter
clockwise (β > 0).
(b) Origin is an unstable focus (α > 0) with trajectories oriented clock
wise (β < 0).
(c) Origin is a stable focus (α < 0) with trajectories oriented clockwise
(β < 0).
(d) Origin is an unstable focus (α > 0) with trajectories oriented counter
clockwise (β > 0).
(e) Origin is a center (α = 0) with trajectories oriented counter
clockwise (β > 0).
(f) Origin is a center (α = 0) with trajectories oriented clockwise (β <
0).
54 phase portrai ts and planar systems
Eigenvectors w must satisfy w
1
−2iw
2
= 0, so we set w
1
= 2iw
2
and treat w
2
as
a free variable. Next, express the eigenvector as a sum of its real and imaginary
parts:
_
w
1
w
2
_
=
_
2iw
2
w
2
_
= w
2
_
2i
1
_
= w
2
__
0
1
_
+ i
_
2
0
__
.
Form the matrix
P =
_
2 0
0 1
_
,
which contains the imaginary part of w in its ﬁrst column and the real part of w
in its second column. Clearly
P
−1
=
_
1
2
0
0 1
_
,
and the real canonical form for A is given by
M = P
−1
AP =
_
0 −2
2 0
_
.
Routine calculation yields the matrix exponential
e
tA
= Pe
tM
P
−1
= P
_
cos 2t −sin2t
sin2t cos 2t
_
P
−1
=
_
cos 2t −2 sin2t
1
2
sin2t cos 2t
_
,
from which it follows that the general solution of our system is
_
x
1
(t)
x
2
(t)
_
= e
tA
_
c
1
c
2
_
=
_
c
1
cos 2t −2c
2
sin2t
1
2
c
1
sin2t + c
2
cos 2t
_
.
The parameter t can be eliminated by algebra, leading to the relationship
x
2
1
+4x
2
2
= c
2
1
+4c
2
2
,
where the right hand side is an arbitrary nonnegative constant. You should
recognize this as the equation of an ellipse. The phase portrait is sketched in
Figure 2.9, and we see that the origin is a center. One way to determine the
orientation of the trajectories is to pick a few convenient points in the phase
li near, constant coeffi ci ent systems 55
x
1
x
2
Figure 2.9. Phase portrait showing the elliptic trajectories x
2
1
+ 4x
2
2
=
constant.
plane and sketch the associated “slope ﬁeld” vectors. For example, suppose we
start from the point (x, y) = (1, 0) on the xaxis in the phase plane. To determine
the direction of motion from that point, we multiply the coefﬁcient matrix A
by (1, 0), obtaining (0, 1). The vector (0, 1), which points straight upward, is
tangent to the solution trajectory passing through (1, 0). This indicates that
trajectories are oriented counterclockwise, as shown in Figure 2.9. Notice that
our phase portrait is simply a “skewed” version of our canonical example of a
center equilibrium (which had circular trajectories).
Example 2.2.4. Sketch the phase portrait for x
/
= Ax, where
A =
_
2 −3
1 −2
_
.
The characteristic equation is λ
2
−1 = 0, so we have real eigenvalues λ = ±1
with different sign. We immediately conclude that the origin in our phase portrait
will be a saddle. You should show that the eigenvalues λ = 1 and λ = −1 give
rise to eigenvectors
_
3
1
_
and
_
1
1
_
,
respectively. By Proposition 2.1.21, the general solution is
_
x
1
x
2
_
= c
1
e
t
_
3
1
_
+ c
2
e
−t
_
1
1
_
.
56 phase portrai ts and planar systems
x
2
x
1
Figure 2.10. Phase portrait showing the skewed saddle The separatrices
are the lines spanned by the eigenvectors [1, 1] and [3, 1].
Consequence: If we set c
1
= 0 and plot the trajectories
c
2
e
−t
_
1
1
_
in the phase plane, we ﬁnd that this is a parametrization of a line of slope 1 as t
varies from −∞ to ∞. The value of c
2
merely selects our “starting point” as we
traverse the line. Similarly, if we had set c
2
= 0, we would have obtained the
parametrization of a line of slope
1
3
. In other words the lines spanned by the two
eigenvectors of A form the separatrices in our phase portrait. This tells us precisely
how to “skew” our canonical example of a saddle equilibrium to obtain a sketch
of the phase portrait, which is shown in Figure 2.10. Notice that the motion
along the separatrix corresponding to the negative eigenvalue is directed inward
to the origin, while the motion along the separatrix corresponding to the positive
eigenvalue is directed outward from the origin.
Before continuing our investigation of the “geometry” associated with planar,
constantcoefﬁcient systems, we follow up our earlier remark that the origin is
always an equilibrium of the linear, homogeneous constantcoefﬁcient system
x
/
= Ax. This admittedly narrow class of systems is not quite as exclusive as
it may appear. Having an equilibrium at the origin is not a severe restriction
at all, because equilibria can always be relocated via a change of variables.
For example, consider the constantcoefﬁcient system x
/
= Ax − b, where b
is a constant vector. Assuming that A is an invertible matrix (i.e., λ = 0 is
li near, constant coeffi ci ent systems 57
not an eigenvalue), we may solve for equilibrium solutions by setting x
/
= 0.
Equivalently, any vector x satisfying Ax = b is an equilibrium. Let x
∗
= A
−1
b
denote the equilibrium solution. Then our original system of odes can be written
as x
/
= Ax − Ax
∗
= A(x −x
∗
). Making the substitution y = x −x
∗
, we obtain a
new system y
/
= Ay. The new constantcoefﬁcient system has its equilibrium
at the origin, and we can solve it using the techniques discussed in previous
sections. We may then recover the solution of our original system by writing
x = y +x
∗
.
The system x
/
= Ax − b is actually a special case of the inhomogeneous
systems we will learn how to solve soon.
2.3. Stable, Unstable, and Center Subspaces
In the previous subsection, we used the words ‘stable’ and ‘unstable’ without
giving a careful deﬁnition of what those terms mean. We know that for homo
geneous constantcoefﬁcient systems x
/
= Ax, the stability of the equilibrium at
the origin is somehow determined by the eigenvalues of A. The eigenvectors of
A determine how the phase portrait is “skewed” from one of the three canonical
phase portraits.
We will give a rigorous deﬁnition of stability of equilibria in the next chapter;
for now, a loose deﬁnition will sufﬁce. An equilibrium solution x
∗
of a system of
odes is stable if, whenever we start from initial conditions that are appropriately
“close” to x
∗
, the resulting solution trajectory never strays too far from x
∗
. A
stable equilibrium is called asymptotically stable if the solution trajectory actually
approaches x
∗
as t →∞. If an equilibrium is not stable, it is called unstable. This
means that there exist initial conditions arbitrarily “close” to x
∗
for which the
solution trajectory is repelled from x
∗
.
For example, if the origin is a center, then we would say the origin is a stable
equilibrium but is not asymptotically stable. Saddle equilibria are unstable,
because it is always possible to choose initial conditions arbitrarily close to the
equilibrium for which the direction of motion in the phase portrait is away from
the equilibrium as t increases. (Just choose any initial conditions not lying on the
separatrix that is oriented towards the equilibrium.)
Saddles are interesting in that there are “special” trajectories in the phase
portrait on which the direction of motion is directed toward the unstable equilib
58 stable, unstable, and center subspaces
rium. In our canonical example of a saddle, one separatrix was oriented toward
the origin, and the other separatrix was oriented away from the origin. The
notion that our underlying space can be decomposed into stable and unstable
“directions” is the subject of our discussion below. First, we recall a familiar
deﬁnition from linear algebra.
Deﬁnition 2.3.1. Let v
1
, v
2
, . . . , v
k
be vectors in R
n
. The span of these vectors is
the set of all linear combinations
c
1
v
1
+ c
2
v
2
+ · · · + c
k
v
k
,
where c
1
, c
2
, . . . c
k
are real numbers.
Notice that if these vectors are linearly independent, then their span forms a
kdimensional subspace of R
n
.
Example 2.3.2. The span of the vectors
_
¸
_
1
0
0
_
¸
_
and
_
¸
_
1
1
0
_
¸
_
in R
3
is the xyplane.
Now consider the homogeneous system x
/
= Ax of odes, where A is an n n
matrix. Let λ
1
, λ
2
, . . . λ
n
denote the eigenvalues of A, repeated according to
algebraic multiplicity. Each eigenvalue can be written in the form λ
j
= α
j
+iβ
j
,
where α
j
and β
j
are real. (Of course, β
j
= 0 if the eigenvalue λ
j
is real.)
Associated with each eigenvalue is a set of eigenvectors (and possibly generalized
eigenvectors).
Deﬁnition 2.3.3. The stable subspace of the system x
/
= Ax is the span of all
eigenvectors and generalized eigenvectors associated with eigenvalues having
negative real part (α
j
< 0). The unstable subspace of the system x
/
= Ax is the span
of all eigenvectors and generalized eigenvectors associated with eigenvalues
having positive real part (α
j
> 0). The center subspace of the system x
/
= Ax is the
span of all eigenvectors and generalized eigenvectors associated with eigenvalues
having zero real part (α
j
= 0).
li near, constant coeffi ci ent systems 59
Notation: The stable, unstable, and center subspaces are denoted by E
s
, E
u
, and
E
c
, respectively.
Example 2.3.4. Consider the system
x
/
=
_
1 2
5 4
_
x.
Letting A denote the coefﬁcient matrix, the characteristic equation is given by
λ
2
−5λ −6 = 0.
Factoring the characteristic equation as (λ−6)(λ+1) = 0, we obtain eigenvalues
λ = −1 and λ = 6. The roots are real and have opposite sign, indicating that the
origin is a saddle. You can verify that
_
−1
1
_
and
_
2
5
1
_
are eigenvectors corresponding to λ = −1 and λ = 6, respectively. By the above
deﬁnition, the stable subspace is given by
E
s
= span
__
−1
1
__
and the unstable subspace is given by
E
u
= span
__
2
5
1
__
.
The center subspace E
c
consists only of the zero vector. Notice that E
s
and E
u
really are subspaces of R
2
, the underlying space. They are straight lines through
the origin and, in this example of a saddle, they correspond to the separatrices
(see Figure 2.11).
We remark that the general solution of this system is given by
_
x
1
(t)
x
2
(t)
_
= c
1
e
−t
_
−1
1
_
+ c
2
e
6t
_
2
5
1
_
.
60 stable, unstable, and center subspaces
x
2
x
1
E
s
E
u
Figure 2.11. Phase portrait showing the stable and unstable subspaces
E
s
and E
u
.
If we start from (nonzero) initial conditions inside the stable subspace (c
1
,= 0
and c
2
= 0), then our solution trajectory will remain in the stable subspace for
all time t, and we will approach the origin as t →∞. Likewise, if we start from
initial conditions inside the unstable subspace, (c
1
= 0 and c
2
,= 0), then our
solution trajectory remains in the unstable subspace for all time t but we always
move away from the origin.
Example 2.3.5. The system
x
/
=
_
1 −8
8 1
_
x
has a coefﬁcient matrix that is already in real canonical form, and its eigenvalues
are 1 ± 8i. Since both eigenvalues have positive real part, E
u
= R
2
whereas
both E
s
and E
c
consist only of the zero vector. Likewise, the system
x
/
=
_
0 −8
8 0
_
x
has a coefﬁcient matrix with eigenvalues ±8i, both of which have zero real part.
In this case, E
c
= R
2
while the stable and unstable subspaces consist only of
the zero vector.
The notions of stable, unstable, and center subspaces are not restricted to
planar systems, as we illustrate in the following example.
li near, constant coeffi ci ent systems 61
Example 2.3.6. Consider the system
x
/
=
_
¸
_
−2 −1 0
1 −2 0
0 0 8
_
¸
_
x,
whose coefﬁcient matrix has a convenient blockdiagonal structure. The eigenval
ues are −2 ±i and 8. The eigenvalue −2 +i gives rise to a complex eigenvector
_
¸
_
0
1
0
_
¸
_
+ i
_
¸
_
1
0
0
_
¸
_
,
and the real eigenvalue has
_
¸
_
0
0
1
_
¸
_
as an eigenvector. Since −2 +i has negative real part, we conclude that the stable
subspace is
E
s
= span
_
¸
_
¸
_
_
¸
_
0
1
0
_
¸
_
,
_
¸
_
1
0
0
_
¸
_
_
¸
_
¸
_
.
Graphically, this is the xyplane in R
3
. The unstable subspace
E
u
= span
_
¸
_
¸
_
_
¸
_
0
0
1
_
¸
_
_
¸
_
¸
_
is onedimensional and corresponds to the zaxis in R
3
. The center subspace E
c
consists only of the zero vector.
Several observations will help us sketch the phase portrait. Any trajectory
starting from nonzero initial conditions in E
s
(the xyplane) will spiral inward
toward the origin while always remaining within E
s
. Any nonzero trajectory
starting in E
u
(the zaxis) will remain within E
u
for all time t and will be oriented
outward from the origin. Finally, any trajectory starting outside E
s
and E
u
will
spiral away from the xyplane but will draw closer to the zaxis as t advances. A
62 stable, unstable, and center subspaces
E
s
E
u
Figure 2.12. Phase portrait of a 3d system with a twodimensional
stable subspace E
s
and a onedimensional unstable subspace E
u
.
sketch of the phase portrait appears in Figure 2.12. As another illustration, the
system
x
/
=
_
¸
_
0 −1 0
1 0 0
0 0 8
_
¸
_
x,
has a coefﬁcient matrix with eigenvalues ±i and 8. The corresponding eigen
vectors are exactly as in the previous example. The only difference is that the
xyplane is now the center subspace because the eigenvectors that span that
plane correspond to an eigenvalue with zero real part. The zaxis is the unstable
subspace. Any trajectory starting from nonzero initial conditions in E
c
(the
xyplane) will circle around the origin, remaining in E
c
for all time t without
being attracted or repelled by the origin. Any nonzero trajectory starting in E
u
(the zaxis) will remain within E
u
for all time t and will be oriented outward
from the origin. Finally, any trajectory starting outside E
c
and E
u
will spiral
away from the xyplane but will always maintain a constant distance from the
zaxis as t advances. In other words, such trajectories are conﬁned to “inﬁnite
cylinders.” A sketch of this phase portrait appears in Figure 2.13.
Before moving on, we introduce some important terminology related to our
above discussion. As usual, consider the system x
/
= Ax where A is an n n
constant matrix. Given an initial condition x
0
= x(0), we know that the unique
solution of this initial value problem is given by x(t) = e
tA
x
0
. In terms of
the phase portrait, once we pick the point x
0
, premultiplying by the matrix
li near, constant coeffi ci ent systems 63
E
c
E
u
Figure 2.13. Phase portrait of a 3d system with a twodimensional
center subspace E
c
and onedimensional unstable subspace E
u
.
e
tA
and letting t increase will trace out a curve in R
n
. Let φ
t
be the function
which associates each different initial condition with its solution trajectory:
φ
t
(x
0
) = e
tA
x
0
.
Deﬁnition 2.3.7. The set of functions φ
t
= e
tA
is called the ﬂow of the system
x
/
= Ax of odes.
The reason for using the name “ﬂow” is that φ
t
describes the motion along
trajectories in the phase space starting from various choices of initial conditions
x
0
.
Deﬁnition 2.3.8. If all eigenvalues of A have nonzero real part, the ﬂow is called
a hyperbolic ﬂow the system x
/
= Ax is called a hyperbolic system, and the origin is
called a hyperbolic equilibrium point.
Example 2.3.9. The system x
/
= Ax where
A =
_
0 1
−9 0
_
,
is nonhyperbolic because the eigenvalues of A are λ = ±3i, both of which have
zero real part. The origin, a center in this case, is a nonhyperbolic equilibrium.
Note that, if an equilibrium is hyperbolic, then the dimension of its center
subspace is zero.
Our next comments concern properties of the stable, unstable, and center
subspaces associated with the system x
/
= Ax. In the examples we gave, there
64 stable, unstable, and center subspaces
are several common themes. First, the pairwise intersections of E
s
, E
u
, and E
c
consist only of the zero vector. Second, the sum of the dimensions of these
three subspaces is always equal to the dimension of the underlying space R
n
.
Finally, if we start from an initial condition x
0
which lies inside one of these
three subspaces, then φ
t
(x
0
) remains within that subspace for all real t. These
observations are now stated formally.
Recall that if S
1
and S
2
are subspaces of a vector space V, then the sum of the
subspaces is deﬁned as
S
1
+ S
2
= {x +y : x ∈ S
1
and y ∈ S
2
} .
The sum S
1
+S
2
is itself a subspace of V, and the concept of sums of vector spaces
is easily extended to larger ﬁnite sums. In the special case where S
1
+ S
2
= V
and S
1
∩ S
2
consists only of the zero vector, we refer to the sum as a direct sum
and write S
1
⊕S
2
= V.
Theorem 2.3.10. Consider the system x
/
= Ax, where A is an n n constant
matrix, and let E
s
, E
u
, and E
c
denote the stable, unstable, and center subspaces
associated with the equilibrium at the origin. Then R
n
= E
s
⊕E
u
⊕E
c
.
Theorem 2.3.11. The subspaces E
s
, E
u
and E
c
are invariant with respect to the ﬂow
φ
t
= e
tA
in the following sense: If x
0
is any initial condition in E
s
, then e
tA
x
0
is
in E
s
for all t. Similar statements hold for E
c
and E
u
.
In the phase portrait, if we start from within one of these three invariant
subspaces, our solution trajectory will never escape from that subspace.
The ﬁnal theorem in this section is one of the most important qualitative results
for the constantcoefﬁcient systems we have studied up to now. It tells us what
sort of behavior we can expect if we know the eigenvalues of the coefﬁcient
matrix A.
Theorem 2.3.12. Consider the linear system x
/
= Ax, where A is an n n
constant matrix. Then the following statements are equivalent:
Given any x
0
∈ R
n
, we have
lim
t→∞
e
tA
x
0
= 0,
and for any nonzero x
0
∈ R
n
, the distance from e
tA
x
0
to the origin tends to ∞
as t → −∞.
li near, constant coeffi ci ent systems 65
All eigenvalues of A have negative real part.
The stable subspace E
s
at the origin is the entire space R
n
.
In other words, if all eigenvalues have negative real part, then the ﬂow directs
us towards the origin as t →∞ and we are repelled from the origin as t → −∞.
The proofs of the three preceding theorems are not difﬁcult, but are omitted.
For linear, homogeneous, constantcoefﬁcient systems, it is easy to classify the
stability of the origin. Namely, the origin is
Asymptotically stable if all eigenvalues of A have negative real part.
Stable if none of the eigenvalues has positive real part.
Unstable if any of the eigenvalues has positive real part.
Example 2.3.13. If
A =
_
1 0
0 −1
_
,
then the origin is unstable because one of the eigenvalues is positive. If
A =
_
0 1
−1 0
_
,
then the origin is stable but not asymptotically stable. If
A =
_
−1 0
0 −6
_
,
then the origin is asymptotically stable.
2.4. Trace and Determinant
A more compact way of classifying phase portraits of planar systems can be
stated in terms of the trace and determinant of the coefﬁcient matrix. Recall the
following
Deﬁnition 2.4.1. The trace of a square matrix A is the sum of the entries on its
main diagonal and is denoted by trA.
66 trace and determi nant
The eigenvalues of a 2 2 matrix A can be expressed in terms of trA and
det A. Suppose
A =
_
a b
c d
_
.
Then the matrix
A −λI =
_
a −λ b
c d −λ
_
has determinant (a −λ)(d −λ) −bc. Equivalently, the characteristic equation is
λ
2
−(a + d)λ + (ad −bc) = 0.
Since trA = a + d and det A = ad −bc, the characteristic equation can also be
written as
λ
2
−(trA)λ +det A = 0. (2.12)
The roots of (2.12) are
λ =
trA ±
_
(trA)
2
−4 det A
2
. (2.13)
Note that the sum of these eigenvalues is trA. This is true of all square matrices,
not just 2 2 matrices (see exercises). We also know that det A is the product
of the eigenvalues. For 2 2 matrices, we have either two real eigenvalues or a
complex conjugate pair of eigenvalues, α ± βi. Thus, for planar systems x
/
= Ax,
we can use trA and det A to classify the origin as a saddle, unstable node, stable
node, unstable focus, stable focus or center:
Case 1: If det A < 0, we claim that the origin is a saddle. To see why, we
must show that the eigenvalues of A are real and have opposite sign. Suppose
indirectly that A has complex conjugate eigenvalues α ± βi. Then the product
of the eigenvalues (which equals det A) would be positive, contradicting our
assumption that det A < 0. It follows that the eigenvalues must be real, and they
must have opposite sign in order for det A < 0. Therefore, the origin is a saddle,
as claimed.
Case 2: Next, suppose that det A > 0 and (trA)
2
−4 det A ≥ 0. From for
mula (2.13), we know that the eigenvalues are real because the discriminant is
positive. Since det A > 0, the eigenvalues have the same sign, and it follows that
li near, constant coeffi ci ent systems 67
the origin is a node. Whether the origin is stable or unstable depends upon trA
(the sum of the eigenvalues):
If trA > 0, then both eigenvalues are positive and the node is unstable.
If trA < 0, then both eigenvalues are negative and the node is stable.
Case 3: Finally, suppose that det A > 0 and (trA)
2
−4 det A < 0. The discrim
inant in formula (2.13) is negative, implying that the eigenvalues are complex
conjugate. The origin is either a focus or a center depending upon the trace of A.
The sum of the eigenvalues α ± βi is 2α, or equivalently trA = 2α.
If trA > 0, the real part of the eigenvalues is positive and the origin is an
unstable focus.
If trA < 0, the real part of the eigenvalues is negative and the origin is a
stable focus.
If trA = 0, the real part of the eigenvalues is zero and the origin is a center.
The above observations are useful in classifying equilibria of planar systems.
Example 2.4.2. Consider the system x
/
= Ax where
A =
_
−2 10
−3 1
_
.
Since det A = 28 > 0, the origin is not a saddle. Since (trA)
2
−4 det A =
−111 < 0, the eigenvalues are complex conjugate, and since trA = −1 < 0, the
eigenvalues have negative real part. Therefore, the origin is a stable focus.
Admittedly, for planar systems it is relatively easy to actually compute the
eigenvalues of the coefﬁcient matrix, avoiding the need for the tracedeterminant
classiﬁcation. However, the above formalism can be very useful for systems in
which the coefﬁcient matrix A contains unspeciﬁed parameters. The inequalities
appearing in Cases 1–3 can be used to determine the ranges of parameter values
for which various types of dynamical behavior will be observed. Such issues will
be explored in our chapter on bifurcations.
2.5. Inhomogeneous Systems
We now extend our earlier work to inhomogeneous linear systems of the form
x
/
= Ax +b(t),
68 i nhomogeneous systems
where A is an n n constant matrix and b(t) is a vector of continuous functions.
Note that b(t) is allowed to involve the independent variable t but none of the
dependent variables. To solve such systems, we will extend the variation of
parameters method covered in introductory differential equations courses.
Variation of parameters. Let us brieﬂy review how to solve ﬁrstorder, linear,
homogeneous odes with only one dependent variable. Speciﬁcally, consider the
ode x
/
(t) = a(t)x(t) + b(t), where x(t) is scalarvalued, not vectorvalued
2
. The
ﬁrst step is to write the ode in the form
x
/
(t) −a(t)x(t) = b(t). (2.14)
We cannot simply integrate both sides with respect to t, because this would lead
to an integral equation
x(t) −
_
a(t)x(t) dt =
_
b(t) dt,
which does not help us solve for x(t). Instead, the trick is to multiply Equa
tion (2.14) by an integrating factor
e
−
_
a(t) dt
(2.15)
to obtain
e
−
_
a(t) dt
_
x
/
−a(t)x
¸
= e
−
_
a(t) dt
b(t).
Equivalently,
d
dt
_
e
−
_
a(t) dt
x(t)
_
= e
−
_
a(t) dt
b(t).
Now we are in a position to integrate both sides with respect to t, after which
we may (hopefully) solve for x(t).
Example 2.5.1. Solve the initial value problem
dx
dt
= −
2x
t
+ln t x(1) = 1.
2
For this scalar ode, the coefﬁcient function a(t) is actually allowed to depend upon the in
dependent variable t. Later, when we solve x
/
= Ax +b(t), we must insist that A be a constant
matrix.
li near, constant coeffi ci ent systems 69
Solution: In our above notation, a(t) = −2/t, from which we calculate
_
2
t
dt = 2 ln t.
(We need not include the integration constant here. After reading ahead a few
lines, make sure you can explain why.) Since our initial condition is given at
t = 1, we may drop the absolute value bars. The integrating factor is e
2 ln t
=
e
ln t
2
= t
2
. Multiplying both sides of our ode by the integrating factor gives
t
2
dx
dt
+2tx = t
2
ln t.
Equivalently,
d
dt
_
t
2
x
_
= t
2
ln t,
and integrating both sides yields
t
2
x =
_
t
2
ln t dt.
Integrating by parts,
t
2
x =
t
3
3
ln t −
t
3
9
+ C,
where C is a constant of integration. The initial condition x(1) = 1 can be used
to calculate C = 10/9, which means that the solution of our initial value problem
is
x(t) =
t
3
ln t −
t
9
+
10
9t
2
.
We can extend the variation of parameters technique to inhomogeneous sys
tems of odes of the form x
/
= Ax + b(t), where A is a constant matrix. To
facilitate this process, we ﬁrst state a lemma which tells us how to write our
usual initial value problem as an integral equation. Reformulating odes as
integral equations will help us frequently in subsequent chapters.
Lemma 2.5.2. Consider the initial value problem x
/
= f (x, t) with x(0) = x
0
. If
f is continuous, then this initial value problem can be written as an integral
equation
x(t) = x
0
+
_
t
0
f (x(s), s) ds. (2.16)
Here, s is a “dummy variable” of integration.
70 i nhomogeneous systems
Proof. First, observe that if we set t = 0 in Equation (2.16), then the equation
reduces to x(0) = x
0
, which means the initial condition is satisﬁed. If we
differentiate both sides of (2.16) using the Fundamental Theorem of Calculus,
then we obtain x
/
(t) = f (x(t), t), and we see that the ode is also satisﬁed.
Now consider the initial value problem
x
/
= Ax +b(t) x(0) = x
0
, (2.17)
where A is an n n constant matrix. By analogy with the onevariable problem
above, we rewrite the equation as x
/
− Ax = b(t). If this were a onevariable
problem with A constant, we would use e
−tA
as the integrating factor. We
claim that for the initial value problem (2.17), we can use the matrix e
−tA
as an
integrating factor. Multiplying both sides of our ode system by e
−tA
yields
matrix
¸..¸
e
−tA
_
x
/
− Ax
_
. ¸¸ .
vector
=
matrix
¸..¸
e
−tA
b(t)
.¸¸.
vector
.
Equivalently,
d
dt
_
e
−tA
x
_
= e
−tA
b(t),
and by Lemma 2.5.2 we have
e
−tA
x(t) = e
−tA
x(t)
¸
¸
¸
t=0
+
_
t
0
e
−sA
b(s) ds = Ix(0) +
_
t
0
e
−sA
b(s) ds,
where I is the identity matrix. Multiplying through by e
tA
, we have established
Theorem 2.5.3. The solution of the inhomogeneous initial value problem (2.17)
is given by
x(t) = e
tA
x(0) +e
tA
_
t
0
e
−sA
b(s) ds. (2.18)
When applying formula (2.18), our main challenge is to actually evaluate the
integral of the vectorvalued function e
−sA
b(s).
Example 2.5.4. Solve the initial value problem
x
/
1
= −x
1
+ x
2
+e
−t
x
1
(0) = 1
x
/
2
= −x
1
−x
2
+2e
−t
x
2
(0) = 1.
li near, constant coeffi ci ent systems 71
Solution: In matrix form, the system becomes
_
x
/
1
x
/
2
_
. ¸¸ .
x
/
=
_
−1 1
−1 −1
_
. ¸¸ .
A
_
x
1
x
2
_
. ¸¸ .
x
+
_
e
−t
2e
−t
_
. ¸¸ .
b
x
0
=
_
1
1
_
.
The matrix A is already in real canonical form, and using α = −1 and β = −1 in
Proposition 2.1.26, we ﬁnd that
e
tA
= e
−t
_
cos(−t) −sin(−t)
sin(−t) cos(−t)
_
= e
−t
_
cos t sin t
−sin t cos t
_
.
Here, we have used the fact that cos(−t) = cos t and sin(−t) = −sin t.
Replacing t with −s and using these same trigonometric facts, we have
e
−sA
= e
s
_
cos s −sin s
sin s cos s
_
.
Therefore, the integrand in formula (2.18) is
e
−sA
b(s) = e
s
_
cos s −sin s
sin s cos s
_ _
e
−s
2e
−s
_
=
_
cos s −2 sin s
sin s +2 cos s
_
.
Integrating each component of this vector separately,
_
t
0
e
−sA
b(s) ds =
_
t
0
_
cos s −2 sin s
sin s +2 cos s
_
ds =
_
sin s +2 cos s
−cos s +2 sin s
_¸
¸
¸
¸
¸
t
0
=
_
sin t +2 cos t −2
−cos t +2 sin t +1
_
.
Putting all of this together, Equation (2.18) tells us that the solution of the initial
value problem is
_
x
1
x
2
_
= e
tA
x(0) +e
tA
_
t
0
e
−sA
b(s) ds
= e
−t
_
cos t sin t
−sin t cos t
_ _
1
1
_
+ e
−t
_
cos t sin t
−sin t cos t
_ _
sin t +2 cos t −2
−cos t +2 sin t +1
_
.
72 i nhomogeneous systems
Expanding these products, our overall result is
x
1
(t) = −e
−t
cos t +2e
−t
sin t +2e
−t
x
2
(t) = e
−t
sin t +2e
−t
cos t −e
−t
.
Readers interested in learning more about the types of linear systems of odes
for which exact solution is possible are encouraged to explore other resources
such as the references listed in the bibliography. The special classes of linear
systems we have discussed up to this point will be sufﬁcient for our purposes in
the remainder of this text.
Exercises
1. If λ is an eigenvalue of an invertible matrix A, show that 1/λ is an eigenvalue
of A
−1
. What are the eigenvectors of A
−1
associated with 1/λ?
2. An n n matrix A is called idempotent if A
2
= A. Show that each eigenvalue
of an idempotent matrix is either 0 or 1.
3. Show that if N is a nilpotent matrix (see Deﬁnition 2.1.32), then zero is the
only eigenvalue of N.
4. Give an example of a 3 3 matrix that is diagonalizable but not invertible.
Then, give an example of a 3 3 matrix that is invertible but not diagonaliz
able.
5. Square matrices A and B are called simultaneously diagonalizable if they are
diagonalizable by the same invertible matrix P. Show that if A and B are
simultaneously diagonalizable, then A and B commute. That is, AB = BA.
6. The cosine of an n n square matrix A is deﬁned in terms of the Maclaurin
series representation of the cosine function:
cos(A) =
∞
∑
k=0
(−1)
k
A
2k
(2k)!
= I −
A
2
2!
+
A
4
4!
+ · · ·
(a) Suppose that D = diag{λ
1
, λ
2
, . . . λ
n
} is a diagonal matrix. What is
cos(D)?
(b) Suppose that A is a diagonalizable matrix and A = PDP
−1
where D is
diagonal. Show that cos(A) = P[cos(D)]P
−1
.
li near, constant coeffi ci ent systems 73
(c) Find cos(A) if
A =
_
−3 6
4 −1
_
.
7. Use the Maclaurin series representation of the sine function to deﬁne sin(A),
where A is an n n matrix. Use your deﬁnition to compute (a) the sine
of the 2 2 zero matrix, and (b) the sine of the diagonal matrix D =
diag{π/2, π/2, π/2}.
8. Consider the two matrices
A =
_
1 0
0 0
_
and B =
_
0 1
0 0
_
.
(a) Show that A and B do not commute.
(b) Show that B is not diagonalizable.
(c) Show that e
A+B
,= e
A
e
B
.
9. For each of the following matrices A, compute e
tA
:
(a)
_
2 25/8
2 2
_
(b)
_
0 −1
−1 2
_
(c)
_
2 −1
2 0
_
.
10. Solve the system
dx
1
dt
= 3x
1
+ x
2
dx
2
dt
= x
1
+3x
2
,
with x
1
(0) = −1 and x
2
(0) = 4.
11. Find the general solution of the system
dx
1
dt
= x
1
+2x
2
+3x
3
,
dx
2
dt
= 2x
2
+8x
3
,
dx
3
dt
= 3x
3
.
HINT: Do you really need to exponentiate a matrix here?
12. Find the general solution of the system
dx
1
dt
= x
1
−5x
2
dx
2
dt
= 5x
1
+ x
2
.
74 i nhomogeneous systems
13. Solve the initial value problem
dx
1
dt
= 5x
1
+10x
2
, x
1
(0) = 1
dx
2
dt
= −x
1
+3x
2
, x
2
(0) = 1.
14. Solve the ivp
dx
1
dt
= x
1
−5x
2
dx
2
dt
= 2x
1
+3x
2
,
with x
1
(0) = 1 and x
2
(0) = 1.
15. Find the general solution of the system x
/
= Ax where A is the matrix
A =
_
0 1
−1 2
_
.
16. Find the general solution of the system x
/
= Ax where A is the matrix
A =
_
¸
_
1 0 0
2 1 0
3 2 1
_
¸
_
.
17. Find the general solution of the system x
/
= Ax where A is the matrix
A =
_
¸
_
1 0 0
−1 2 0
1 0 2
_
¸
_
.
18. Solve the initial value problem x
/
= Ax where
A =
_
¸
_
0 1 0
0 0 1
−8 −12 −6
_
¸
_
and x
1
(0) = 1, x
2
(0) = 2, and x
3
(0) = 3.
19. For each of the following matrices A, sketch the phase portrait for the linear
system x
/
= Ax. In each case, indicate whether the origin is a stable node,
li near, constant coeffi ci ent systems 75
unstable node, stable focus, unstable focus, saddle, or center.
A =
_
2 0
0 2
_
A =
_
−1 0
0 3
_
A =
_
1 −3
3 1
_
.
20. For the following matrices A, carefully sketch the phase portrait for the
linear system x
/
= Ax. In each case, identify the stable subspace E
s
and the
unstable subspace E
u
.
A =
_
1 2
4 4
_
A =
_
0 −1
9 0
_
.
Hint: Try to mimic examples that appear in the text.
21. (Trace and determinant.) Without ﬁnding the eigenvalues of the coefﬁcient
matrices of the following systems, determine whether the origin is a stable
node, unstable node, stable focus, unstable focus, or center.
(a) x
/
=
_
−1 1
3 2
_
x (b) x
/
=
_
2 1
1 2
_
x
(c) x
/
=
_
−2 1
1 −1
_
x (d) x
/
=
_
2 −1
5 −2
_
x.
22. The trace of any n n matrix A is equal to the sum of its eigenvalues; the
purpose of this exercise is to prove this statement for diagonalizable matrices A.
(a) Suppose A and B are n n matrices. Show that tr(AB) = tr(BA).
(b) Using the result from Part (a), show that if A and B are similar, then
trA = trB.
(c) Suppose that A is diagonalizable and that A = PDP
−1
where D is diago
nal. Use the result from Part (b) to explain why the trace of A is equal to
the sum of its eigenvalues.
23. Consider the constantcoefﬁcient system x
/
= Ax where
A =
_
¸
_
4 0 −3
0 −2 0
3 0 4
_
¸
_
.
76 i nhomogeneous systems
Determine the stable, unstable and center subspaces E
s
, E
u
and E
c
associated
with the equilibrium at the origin.
24. Consider the system x
/
= Ax, where
A =
_
¸
_
−3 0 0
0 2 −4
0 4 2
_
¸
_
.
Identify E
s
, E
u
and E
c
for this system, and provide a rough sketch of the
phase portrait.
25. Determine E
s
, E
u
and E
c
for the following system, and describe the ﬂow
starting from initial conditions within each of these subspaces.
x
/
=
_
¸
_
0 −3 0
3 0 0
0 0 −1
_
¸
_
x.
26. Each of the following systems contains an unspeciﬁed constant α. For each
system, (i) determine the eigenvalues of the coefﬁcient matrix in terms of α;
(ii) ﬁnd the critical values of α at which the qualitative nature of the phase
portrait experiences a dramatic change; and (iii) sketch the phase portrait for
different choices of α: one just below each critical value, and one just above
each critical value.
(a) x
/
=
_
1 α
2
4 1
_
x (b) x
/
=
_
−1 1
1 α
_
x (c) x
/
=
_
−1 α
1 2
_
x.
27. Solve the initial value problem
dy
dx
= (−tan x)y +sec x, y
_
π
4
_
=
√
2.
28. Solve the initial value problem
dy
dx
= −
y
x
+e
x
2
, y(1) = 3.
li near, constant coeffi ci ent systems 77
29. Solve the initial value problem
x
/
=
_
−1 2
0 1
_
x +
_
1
2t
_
, x
0
=
_
1
0
_
.
30. Solve the initial value problem
dx
dt
= x +2t x(0) = 2
dy
dt
= −y +e
t
y(0) = 1.
31. Solve the initial value problem
dx
1
dt
= −x
1
−2x
2
+cos 2t, x
1
(0) = 3
dx
2
dt
= 2x
1
−x
2
+sin2t, x
2
(0) = 3.
CHAPTER 3
Nonlinear Systems: Local Theory
W
e now turn our attention to nonlinear systems of odes, which present
a host of new challenges. Obtaining exact analytical solutions to such
systems is usually impossible, so we must settle for qualitative descriptions of
the dynamics. On the other hand, nonlinear systems can exhibit a wide variety
of behaviors that linear systems cannot. Moreover, most dynamical phenomena
in nature are inherently nonlinear.
Consider a general system of the form x
/
= f (x, t), where x ∈ R
n
is a vector
of unknowns and
f (x, t) =
_
¸
¸
¸
¸
¸
_
f
1
(x
1
, x
2
, . . . x
n
, t)
f
2
(x
1
, x
2
, . . . x
n
, t)
.
.
.
f
n
(x
1
, x
2
, . . . x
n
, t)
_
¸
¸
¸
¸
¸
_
is a vectorvalued function f : R
n+1
→ R
n
. In what follows, we shall work
only with autonomous systems—those of the form x
/
= f (x) where f : R
n
→R
n
does not explicitly involve the independent variable t. This is actually not a
severe restriction at all, because nonautonomous systems can be converted to
autonomous ones by introducing an extra dependent variable. For example, the
nonautonomous system
_
x
/
1
x
/
2
_
=
_
cos(x
1
t) + x
2
t
2
+ x
2
1
_
78
nonli near systems: local theory 79
becomes autonomous if we introduce a third dependent variable x
3
= t and
corresponding ode x
/
3
= 1. The result is the autonomous system
_
¸
_
x
/
1
x
/
2
x
/
3
_
¸
_
=
_
¸
_
cos(x
1
x
3
) + x
2
x
2
3
+ x
2
1
1
_
¸
_
,
whose order is one larger than the original nonautonomous system. Hence, it
will be sufﬁcient to consider autonomous systems only.
For autonomous systems x
/
= f (x), it is straightforward to determine whether
the system is linear or nonlinear.
Deﬁnition 3.0.5. Consider the system x
/
= f (x), where f : R
n
→ R
n
. The
system of odes is linear if the function f satisﬁes f (αx +y) = αf (x) + f (y) for
all vectors x, y ∈ R
n
and all scalars α ∈ R. Otherwise, the system of odes is
called nonlinear.
Example 3.0.6. The righthand side of the ode
dx
dt
= x
2
is f (x) = x
2
, and
f (x + y) = (x + y)
2
= x
2
+2xy + y
2
,= x
2
+ y
2
= f (x) + f (y).
Therefore, the ode is nonlinear.
Warning: Be careful when classifying odes as autonomous/nonautonomous
or linear/nonlinear. These concepts are independent of one another. Here are
several onedimensional examples to reinforce this point:
dx
dt
= x is linear and autonomous.
dx
dt
= x +cos t is linear and nonautonomous.
dx
dt
= cos x is nonlinear and autonomous.
dx
dt
= t +cos x is nonlinear and nonautonomous.
There are very few analytical techniques for solving nonlinear odes. Separation
of variables is one such method, but it has very limited scope.
Example 3.0.7. Consider the nonlinear, autonomous ode
dx
dt
= x −x
2
with initial
condition x(0) = 1/2. The fact that the ode is autonomous makes separating the
variables easy:
1
x −x
2
dx
dt
= 1.
80
Integrating both sides with respect to t yields
_
1
x −x
2
dx =
_
dt,
and a partial fractions decomposition simpliﬁes the lefthand side as
_
1
x
dx +
_
1
1 −x
dx =
_
dt.
Hence, ln x −ln 1 − x = t + C where C is an integration constant. In the
vicinity of the initial condition x(0) = 1/2, we are dealing with x values for
which both x and 1 −x are positive quantities, allowing us to drop the absolute
value bars. Moreover, the initial condition tells us that C = 0. Solving for x then
gives
x(t) =
e
t
1 +e
t
.
Notice that as t →∞, the solution satisﬁes x → 1 and as t → −∞, we have x → 0.
The linear systems we considered in the previous chapter cannot exhibit this
sort of behavior. In fact, the only systems x
/
= Ax whose nonconstant solutions
remain bounded (i.e., trapped in a ﬁnite interval) for all time t are those for
which all eigenvalues of A have zero real part.
The remarks in the preceding example allude to the fact that nonlinear systems
can exhibit very different behavior from the linear ones we have considered up
to now. The next two examples help drive home this point.
Example 3.0.8. In the previous chapter, we learned that the linear system x
/
=
Ax +b(t) with initial condition x(0) = x
0
has a unique solution given by the
variation of parameters formula (2.18). By contrast, the nonlinear system
dx
dt
= 3x
2/3
x(0) = 0
has at least two solutions. Clearly one solution is the constant function x(t) = 0.
Separation of variables produces a second solution x(t) = t
3
, and therefore this
initial value problem does not have a unique solution.
Example 3.0.9. Consider the nonlinear system
dx
dt
= x
2
x(0) = 1.
nonli near systems: local theory 81
By separation of variables, you can show that x(t) = (1 −t)
−1
is the solution.
Unfortunately, the solution blows up as t → 1
−
. Although the function x(t) =
(1 −t)
−1
is deﬁned for t > 1, it does not make sense to regard this as the solution
of the initial value problem for t ≥ 1 since something went horribly wrong with
our system at t = 1. Note that solutions of the linear equations we studied in the
previous chapter cannot blow up in ﬁnite time.
Moral: For practical purposes, we need a way of determining whether a non
linear system x
/
= f (x) has a unique solution, and whether solutions are well
behaved. In practice, if we obtain multiple solutions to an initial value problem
which is supposed to model some natural phenomenon, then which solution (if
any) is physically relevant? To address the issue of whether a nonlinear system
has a unique solution, we will typically approximate the dynamics using linear
systems.
3.1. Linear Approximations of Functions of Several Variables
One approach toward tackling nonlinear systems x
/
= f (x) involves making
successive linear approximations, a process known as Picard iteration. Before
explaining that procedure, we expand upon some of the ideas covered in multi
variable calculus. The purpose of this section is to explain how to obtain a linear
approximation of a function f : R
n
→R
n
.
For functions f (x) of a single variable, the derivative provides a means for
writing down a linear approximation. Namely, the best linear approximation of
f (x) near a point x = a is given by the tangent line approximation
f (x) ≈ f (a) + f
/
(a) · (x −a).
In multivariable calculus, you learned that graphs of functions f : R
2
→ R
of two variables are surfaces in R
3
. Given a point (x
0
, y
0
) ∈ R
2
, the best linear
approximation of f near that point is given by the tangent plane approximation
f (x, y) ≈ f (x
0
, y
0
)
+
∂ f
∂x
(x
0
, y
0
) · (x −x
0
) +
∂ f
∂y
(x
0
, y
0
) · (y −y
0
).
82 li near approxi mati ons of functi ons of several vari ables
Recall that if f (x, y) is a function of two variables, its gradient ∇f is the vector
∇f (x, y) =
_
∂ f
∂x
,
∂ f
∂y
_
.
Hence, an equivalent way of writing the equation for the tangent plane approxi
mation is
f (x, y) ≈ f (x
0
, y
0
) +∇f (x
0
, y
0
) • (x −x
0
, y −y
0
)
. ¸¸ .
dot product
.
For functions f : R
n
→ R, linear approximation is simply an extension of the
tangent plane approximation. If (x
1
, x
2
, . . . x
n
) is some point in R
n
, then the
linear approximation of f near that point is given by
f (x
1
, x
2
, . . . x
n
) ≈ f (x
1
, x
2
, . . . x
n
) +∇f (x
1
, x
2
, . . . x
n
) • (x
1
−x
1
, x
2
−x
2
, . . . x
n
−x
n
),
where
∇f (x
1
, x
2
, . . . x
n
) =
_
∂ f
∂x
1
,
∂ f
∂x
2
, . . .
∂ f
∂x
n
_
.
We are now ready to explain how to obtain linear approximations for functions
f : R
n
→R
n
. If f is such a function, we can write f in terms of its component
functions
f (x
1
, x
2
, . . . x
n
) =
_
¸
¸
¸
¸
¸
_
f
1
(x
1
, x
2
, . . . x
n
)
f
2
(x
1
, x
2
, . . . x
n
)
.
.
.
f
n
(x
1
, x
2
, . . . x
n
)
_
¸
¸
¸
¸
¸
_
. (3.1)
Since each component function f
i
is a function from R
n
→R, we know how to
write down their linear approximations near a speciﬁc point (x
1
, x
2
, . . . x
n
) ∈ R
n
.
Namely, the right hand side of Equation (3.1) is approximated by
_
¸
¸
¸
¸
¸
_
f
1
(x
1
, x
2
, . . . x
n
) +∇f
1
(x
1
, x
2
, . . . x
n
) • (x
1
−x
1
, x
2
−x
2
, . . . x
n
−x
n
)
f
2
(x
1
, x
2
, . . . x
n
) +∇f
2
(x
1
, x
2
, . . . x
n
) • (x
1
−x
1
, x
2
−x
2
, . . . x
n
−x
n
)
.
.
.
f
n
(x
1
, x
2
, . . . x
n
) +∇f
n
(x
1
, x
2
, . . . x
n
) • (x
1
−x
1
, x
2
−x
2
, . . . x
n
−x
n
)
_
¸
¸
¸
¸
¸
_
. (3.2)
nonli near systems: local theory 83
Finally, we rewrite the vector (3.2) in the equivalent form
_
¸
¸
¸
¸
¸
_
f
1
(x
1
, x
2
, . . . x
n
)
f
2
(x
1
, x
2
, . . . x
n
)
.
.
.
f
n
(x
1
, x
2
, . . . x
n
)
_
¸
¸
¸
¸
¸
_
. ¸¸ .
vector in R
n
+
_
¸
¸
¸
¸
¸
_
∇f
1
(x
1
, x
2
, . . . x
n
)
∇f
2
(x
1
, x
2
, . . . x
n
)
.
.
.
∇f
n
(x
1
, x
2
, . . . x
n
)
_
¸
¸
¸
¸
¸
_
. ¸¸ .
nn matrix
_
¸
¸
¸
¸
¸
_
x
1
−x
1
x
2
−x
2
.
.
.
x
n
−x
n
_
¸
¸
¸
¸
¸
_
. ¸¸ .
vector in R
n
. (3.3)
Here, it is understood that the gradients in (3.3) are written out as row vectors.
Equation (3.3) is the best linear approximation for a function f : R
n
→ R
n
near the point (x
1
, x
2
, . . . x
n
). The matrix appearing in (3.3) will be important
throughout our study of nonlinear odes, and it has a special name.
Deﬁnition 3.1.1. Suppose f : R
n
→R
n
is written in terms of its components as
in (3.1), and further suppose that the ﬁrst partial derivatives of each component
function exist. That is, ∂ f
i
/∂x
j
exists for 1 ≤ i, j ≤ n. Then the Jacobian of f is the
matrix
J f (x
1
, x
2
, . . . x
n
) =
_
¸
¸
¸
¸
¸
¸
_
∂ f
1
∂x
1
∂ f
1
∂x
2
· · ·
∂ f
1
∂x
n
∂ f
2
∂x
1
∂ f
2
∂x
2
· · ·
∂ f
2
∂x
n
.
.
.
.
.
.
.
.
.
.
.
.
∂ f
n
∂x
1
∂ f
n
∂x
2
· · ·
∂ f
n
∂x
n
_
¸
¸
¸
¸
¸
¸
_
. (3.4)
Evaluating a Jacobian matrix J f at a point (x
1
, x
2
, . . . x
n
) ∈ R
n
is perfectly
analogous to evaluating the derivative of a singlevariable function f at a point
x = a. Readers who plan to take advanced calculus/analysis courses will learn
that, if all partial derivatives in a Jacobian matrix are continuous, then the
Jacobian matrix represents the derivative of a function f : R
n
→R
n
.
Example 3.1.2. Consider the function f : R
2
→R
2
deﬁned by
f (x, y) =
_
x
2
+ y
3
x cos(xy)
_
.
Then
J f (x, y) =
_
∂ f
1
/∂x ∂ f
1
/∂y
∂ f
2
/∂x ∂ f
2
/∂y
_
=
_
2x 3y
2
cos(xy) −xy sin(xy) −x
2
sin(xy)
_
.
84 fundamental exi stence and uni queness theorem
In particular,
J f (0, 0) =
_
0 0
1 0
_
.
3.2. Fundamental Existence and Uniqueness Theorem
The notion of Jacobian matrices will become important when we state conditions
which guarantee that a general nonlinear system x
/
= f (x) with x(0) = x
0
has a
unique solution. By Lemma 2.5.2, this initial value problem can be written as an
integral equation
x(t) = x
0
+
_
t
0
f (x(s)) ds.
Picard’s method of successive approximations states that, under reasonable
assumptions on f , the solution of this integral equation can be constructed
recursively by setting u
0
(t) = x
0
and deﬁning the functions
u
k+1
(t) = x
0
+
_
t
0
f (u
k
(s)) ds
for each k ≥ 0. Provided that f is reasonably wellbehaved, the sequence of
functions {u
k
(t)} converges to x(t), the solution of the initial value problem, on
some time interval t ∈ (−α, α). We illustrate the Picard iteration procedure via
an example.
Example 3.2.1. Consider the onedimensional system x
/
= ax with initial condi
tion x(0) = x
0
. In this case, f (x) = ax and we know in advance that the solution
of this problem is simply x = x
0
e
at
. To start the process of Picard iteration, we
deﬁne u
0
(t) = x
0
, a constant. Then
u
1
(t) = x
0
+
_
t
0
f (u
0
(s)) ds = x
0
+
_
t
0
ax
0
ds = x
0
+ ax
0
t = x
0
(1 + at).
The next iterate is
u
2
(t) = x
0
+
_
t
0
f (u
1
(s)) ds = x
0
+
_
t
0
ax
0
(1 + as) ds = x
0
_
1 + at +
(at)
2
2
_
.
nonli near systems: local theory 85
After computing a few more iterates, we are led to conjecture that
u
k
(t) = x
0
k
∑
j=0
(at)
j
j!
,
which can be proved by straightforward induction. Letting k →∞, we ﬁnd that
lim
k→∞
u
k
(t) = x
0
∞
∑
j=0
(at)
j
j!
= x
0
e
at
.
Here, we have recognized the inﬁnite series as the Maclaurin series representation
of the exponential function e
at
. Notice that the limit of our sequence of functions
u
k
(t) is x(t), the solution of the initial value problem. In this example, the
solution is valid for all real t.
We now state a criterion under which the Picard iteration process will actually
produce the unique solution of our initial value problem.
Theorem 3.2.2. (Fundamental Existence & Uniqueness Theorem): Consider
the initial value problem x
/
= f (x) and x(0) = x
0
, where f : R
n
→R
n
. Suppose
that all partial derivatives in the Jacobian matrix J f (x) are continuous for all x in
the vicinity of the initial condition x
0
. Then there exists a positive number α such
that the initial value problem has a unique solution x(t) in the interval [−α, α].
Remark. (i) This Theorem is local. It does not guarantee that a unique solution
will exist for all time t. Conditions for global existence and uniqueness will be
provided later. (ii) It is possible to state a stronger version of this theorem which
does not require that J f (x) consist of continuous functions. See Perko [8] for
details. (iii) There are several ways to prove this theorem (see [7]). The most
intuitive proof requires familiarity with the notions of uniform convergence and
the contraction mapping principle.
Example 3.2.3. Consider the initial value problem
dx
1
dt
= 2x
1
−3x
1
x
2
x
1
(0) = 3
dx
2
dt
= x
1
x
2
−4x
2
x
2
(0) = 8.
86 global exi stence, dependence on i ni ti al condi ti ons
In this case, f : R
2
→R
2
is given by
f (x
1
, x
2
) =
_
f
1
(x
1
, x
2
)
f
2
(x
1
, x
2
)
_
=
_
2x
1
−3x
1
x
2
x
1
x
2
−4x
2
_
.
The corresponding Jacobian matrix is
J f (x
1
, x
2
) =
_
2 −3x
2
−3x
1
x
2
x
1
−4
_
.
Not only are the entries of this matrix continuous at the point
x
0
=
_
3
8
_
,
they are continuous at all points (x
1
, x
2
) ∈ R
2
. Therefore, there exists α > 0 such
that this initial value problem has a unique solution on the interval [−α, α].
Example 3.2.4. Recall the initial value problem
dx
dt
= 3x
2/3
, x(0) = 0,
for which we found two solutions: x(t) = 0 and x(t) = t
3
. In this one
dimensional example, we have f (x) = 3x
2/3
, and the “Jacobian” of f is simply
its usual derivative. That is, J f (x) = f
/
(x) = 2x
−1/3
. Notice that f
/
(x) is NOT
continuous at x = 0, which is precisely the value of x assigned by the initial con
dition. Hence, we have not satisﬁed the conditions of the Fundamental Existence
and Uniqueness Theorem. If the initial condition had been x(0) = 1, then our
problem would have a unique solution. (Warning: Simply failing to satisfy the
conditions of Theorem 3.2.2 is not enough to logically conclude that the initial
value problem does not have a unique solution. It merely tells us that we cannot
immediately conclude that our problem does have a unique solution.)
3.3. Global Existence, Dependence on Initial Conditions
Before jumping into a discussion of qualitative theory of nonlinear systems, we
address two questions which follow upon the theory in the previous section.
If we slightly change our initial conditions, can this make a profound impact
nonli near systems: local theory 87
on solutions of our initial value problem? Under what circumstances can we
guarantee global existence of solutions as opposed to the local existence that
Theorem 3.2.2 ensures? We will answer these questions after recalling some basic
facts from singlevariable calculus.
Lemma 3.3.1. (Extreme Value Theorem). If a function f (x) is continuous on
a closed interval [a, b], then f (x) achieves both a minimum value m and a
maximum value M on this interval.
For example, on the closed interval [0, 3π/4], the function f (x) = sin x achieves
a maximum value of 1 at x = π/2 and a minimum value of 0 at x = 0. By contrast,
on the open interval (0, 1), the function g(x) = x
2
attains neither a maximum
nor a minimum value.
Lemma 3.3.2. (Mean Value Theorem). Suppose f (x) is continuous on [a, b]
and differentiable on (a, b). Then there exists at least one point c ∈ (a, b) such
that
f
/
(c) =
f (b) − f (a)
b −a
.
Geometrically, the mean value theorem is easy to explain. Refer to Figure 3.1,
which shows a graph of a differentiable function f (x) on an interval [a, b]. The
dashed line has slope
f (b) − f (a)
b −a
.
In the ﬁgure, there happen to be two x values in the interval (a, b) at which the
slope of the tangent line (solid line segments) is exactly equal to the slope of the
dashed line. The mean value theorem guarantees that at least one such x value
will always exist.
Lemma 3.3.3. Suppose that f (x) is continuously differentiable on a closed inter
val [a, b]. (That is, f
/
(x) exists and is continuous on that interval.) Then there
exists a positive number K such that  f (x) − f (y) ≤ Kx −y for all x, y ∈ [a, b].
Proof. Since f
/
(x) is continuous on the closed interval [a, b], Lemma 3.3.1 tells
us that f
/
(x) achieves both a minimum m and a maximum M somewhere on
this interval. Let K = max {m, M}. Suppose x, y ∈ [a, b] and, without loss
of generality, x < y. The the mean value theorem 3.3.2 says that there exists a
88 global exi stence, dependence on i ni ti al condi ti ons
f(x)
x
Figure 3.1. Illustration of the meanvalue theorem. See text for details.
number c ∈ (x, y) such that
f
/
(c) =
f (y) − f (x)
y −x
.
Taking absolute values, we ﬁnd that
¸
¸
¸
¸
f (y) − f (x)
y −x
¸
¸
¸
¸
=  f
/
(c) ≤ K.
Therefore,  f (y) − f (x) ≤ Ky −x regardless of x and y.
Lemma 3.3.4. Suppose f is continuous on the interval [a, b]. Then
¸
¸
¸
¸
_
b
a
f (x) dx
¸
¸
¸
¸
≤
_
b
a
 f (x) dx.
Proof. Exercise.
Lemma 3.3.5. (Gronwall’s Inequality). Suppose f is nonnegative and continu
ous on R, and suppose there exist positive constants C and K such that
f (t) ≤ C + K
_
t
0
f (s) ds.
for all t ∈ [0, a]. Then f (t) ≤ Ce
Kt
for all t ∈ [0, a].
Proof. Exercise.
nonli near systems: local theory 89
Our ﬁnal Lemma will reference a deﬁnition that the reader may be familiar
with.
Deﬁnition 3.3.6. If x is a vector in R
n
, then the Euclidean norm of x is deﬁned as
x
2
=
_
x
2
1
+ x
2
2
+ · · · + x
2
n
.
If x and y are two points in R
n
, the Euclidean distance between these two points
is given by
x −y
2
=
_
(x
1
−y
1
)
2
+ (x
2
−y
2
)
2
+ · · · + (x
n
−y
n
)
2
.
The Euclidean norm of a vector is a generalization of our usual notion of the
length of a vector, and Euclidean distance is a generalization of our usual notion
of distance in two or three dimensions.
Lemma 3.3.7. (Triangle Inequality). If x, y ∈ R
n
, then x +y
2
≤ x
2
+y
2
.
In the special case n = 1, this means x + y ≤ x + y.
We now use the above lemmas to prove an important theorem which tells us
how changing the initial condition may affect the solution of an initial value
problem.
Theorem 3.3.8. (Dependence on Initial Conditions). Suppose f is continu
ously differentiable on R and consider the two initial value problems
x
/
= f (x)x(0) = x
0
y
/
= f (y)y(0) = y
0
= x
0
+ C.
Then the solutions of these two problems separate at most exponentially fast as
t increases (at least over the time interval in which the solutions exist). More
exactly, there exists a positive constant K such that x(t) −y(t) ≤ Ce
Kt
.
Proof. Notice that x and y satisfy the same ode, but with different initial condi
tions. We write both of these initial value problems as integral equations
x(t) = x
0
+
_
t
0
f (x(s)) ds
y(t) = y
0
+
_
t
0
f (y(s)) ds = x
0
+ C +
_
t
0
f (y(s)) ds.
90 global exi stence, dependence on i ni ti al condi ti ons
Measuring the gap between the solutions,
y(t) −x(t) =
¸
¸
¸
¸
x
0
+ C +
_
t
0
f (y(s)) ds −
_
x
0
+
_
t
0
f (x(s)) ds
_¸
¸
¸
¸
=
¸
¸
¸
¸
C +
_
t
0
f (y(s)) − f (x(s)) ds
¸
¸
¸
¸
.
By Lemma 3.3.7, we have
y(t) −x(t) ≤ C +
¸
¸
¸
¸
_
t
0
f (y(s)) − f (x(s)) ds
¸
¸
¸
¸
,
from which Lemma 3.3.4 tells us that
y(t) −x(t) ≤ C +
_
t
0
 f (y(s)) − f (x(s)) ds.
Since f is continuously differentiable, Lemma 3.3.3 says that there exists a positive
constant K such that  f (y(s)) − f (x(s)) ≤ Ky(s) − x(s) which provides the
estimate
y(t) −x(t) ≤ C + K
_
t
0
y(s) −x(s) ds.
Finally, Gronwall’s Inequality (Lemma 3.3.5) gives
y(t) −x(t) ≤ Ce
Kt
.
This means that solutions of the two initial value problems we started with can
separate at most exponentially fast.
With Theorem 3.3.8 in mind, we give an important deﬁnition.
Deﬁnition 3.3.9. An initial value problem
x
/
= f (x) x(0) = x
0
is called wellposed if each of the following criteria is satisﬁed.
Existence: The problem has at least one solution.
Uniqueness: The problem has at most one solution.
Dependence on Initial Conditions: A slight change in initial conditions does not
profoundly impact the solution. In other words, solutions of “nearby” initial
value problems do not separate faster than exponentially.
nonli near systems: local theory 91
In practice, one always hopes to work with wellposed problems. Non
existence of a solution has obvious negative implications, and typically raises
questions as to how best to create an “approximate” solution. Nonuniqueness of
solutions can sometimes be remedied by imposing additional requirements that
solutions must satisfy (e.g., boundary conditions or initial conditions) allowing
us to single out one particular solution of interest. Sensitive dependence on initial
conditions can be a severe problem. If the behavior of a system is extremely
sensitive to the choice of initial conditions, then the solutions may be completely
unreliable in making predictions of future behavior. Mathematical models of the
weather tend to suffer from this latter drawback, making it difﬁcult to forecast
beyond a few days ahead.
We now turn our attention to the question of global existence of solutions
of the initial value problem x
/
= f (x) and x(0) = x
0
, where f is continuously
differentiable. The Fundamental Existence and Uniqueness Theorem 3.2.2 and
Theorem 3.3.8 guarantee that this problem is wellposed, but they only ensure
that the solution will exist locally (i.e., in some interval (α, β) containing t = 0).
Lemma 3.3.10. Consider the initial value problem
x
/
= f (x) x(0) = x
0
where f is continuously differentiable on R
n
. Then there is a maximal interval
J = (α, β) over which the initial value problem has a unique solution x(t). This
maximal interval of existence J is an open interval containing t = 0.
Proof. See Section 2.4 of Perko [8].
Remark. (i) Let us clarify what we mean by the maximal interval of existence.
Lemma 3.3.10 states that if the initial value problem has a solution y(t) on
some interval I, then I must be a subset of J, and y(t) = x(t) for all t ∈ I. (ii)
If the maximal interval of existence if (−∞, ∞), then we say that the solution
exists globally. (iii) The statement of Lemma 3.3.10 need not require that f be
continuously differentiable on all of R
n
, only on some open set containing the
initial condition x
0
. Our less general formulation of 3.3.10 will facilitate the
statement and proof of the next lemma, as well as our main Theorem (see 3.3.13
below).
92 global exi stence, dependence on i ni ti al condi ti ons
Lemma 3.3.11. Suppose that the initial value problem in Lemma 3.3.10 has a
maximum interval of existence (α, β) where β < ∞. Then given any positive
number M (no matter how large), there exists some t ∈ (α, β) such that x(t)
2
>
M.
Proof. See Section 2.4 of Perko [8].
In words, Lemma 3.3.11 tells us that if the solution of the initial value problem
fails to exist for all positive time t, then the solution curve cannot possibly
remain conﬁned within some ﬁxed distance M of the origin throughout the
entire interval of existence.
Lemma 3.3.12. (Strong Gronwall Inequality). Suppose g ≥ 0 is continuous on
R and that there exist positive constants C, B, and K such that
g(t) ≤ C + Bt + K
_
t
0
g(s) ds
for t ≥ 0. Then
g(t) ≤ Ce
Kt
+
B
K
_
e
Kt
−1
_
.
Proof. Straightforward extension of the proof of the basic Gronwall inequal
ity 3.3.5.
Theorem 3.3.13. (Global Existence). Consider the initial value problem
x
/
= f (x) x(0) = x
0
,
where f : R
n
→R
n
is continuously differentiable. If there exist positive constants
K and B such that f (x)
2
≤ Kx
2
+ B for all x ∈ R
n
, then the initial value
problem has a unique solution valid for all real t.
Remark. The condition f (x)
2
≤ Kx
2
+ B essentially means that the “size”
of the vector f (x)
2
grows no faster than linearly as x moves away from the
origin.
Proof. We give the proof for 0 < t < ∞, arguing by contradiction. Suppose the
maximal interval of existence is (α, β) where β < ∞. Let g(t) = x
2
, where
x = x(t) denotes the solution of the initial value problem. Then g is continuous,
nonli near systems: local theory 93
but not necessarily differentiable. Since
x(t) = x
0
+
_
t
0
f (x(s)) ds,
we make the estimate
g(t) =
_
_
_
_
x
0
+
_
t
0
f (x(s)) ds
_
_
_
_
2
≤ x
0

2
+
_
_
_
_
_
t
0
f (x(s)) ds
_
_
_
_
2
≤ x
0

2
+
_
t
0
f (x(s))
2
ds ≤ x
0

2
+
_
t
0
K x(s)
2
+ B ds
= x
0

2
+
_
t
0
B ds + K
_
t
0
x(s)
2
ds = x
0

2
+ Bt + K
_
t
0
g(s) ds.
By the Strong Gronwall Inequality 3.3.12, we conclude that
g(t) ≤ x
0

2
e
Kt
+
B
K
_
e
Kt
−1
_
.
This means that x(t)
2
can grow at most exponentially fast. Returning to our
indirect assumption that β < ∞, Lemma 3.3.11 tells us that our solution x(t)
will never remain conﬁned to a given bounded region throughout its maximal
interval of existence. In particular, suppose we choose
M = x
0

2
e
βK
+
B
K
_
e
βK
−1
_
in the statement of Lemma 3.3.11. Then the Lemma implies that
g(t) = x(t)
2
> M
for some t ∈ (α, β). This contradicts our above inequality for g(t). Therefore,
β = ∞ and the solution of the initial value problem exists for all positive t.
Extending this proof to establish existence for all negative t is straightforward.
Example 3.3.14. The right hand side of the planar system
dx
dt
= f
1
(x, y) = sin(x
2
+ y
2
)
dx
dt
= f
2
(x, y) = cos(x
2
+ y
2
)
94 equi li bri a and li neari zati on
consists of functions that are continuously differentiable. Since the sine and
cosine functions remain bounded between 1 and 1, we have f (x, y)
2
≤
√
1
2
+1
2
=
√
2. Therefore, f obeys an estimate of the form f 
2
≤ Kx
2
+ B,
where K = 0 and B =
√
2. The global existence theorem 3.3.13 guarantees that
any solution of this system of odes will exist for all time t.
3.4. Equilibria and Linearization
When approximating nonlinear systems with linear ones, one typically works in
the vicinity of equilibrium solutions (for reasons that we shall soon reveal). We
begin this section with some deﬁnitions which will be used frequently throughout
the remainder of this chapter.
Deﬁnition 3.4.1. An equilibrium solution of x
/
= f (x) is any constant vector x
∗
such that f (x
∗
) = 0.
Deﬁnition 3.4.2. Let be a ﬁxed, positive number and suppose x ∈ R
n
. The
open ball of radius centered at x is the set of all points whose distance from x is
less than . We will use the notation
B(x, ) = {y ∈ R
n
such that x −y
2
< } .
Deﬁnition 3.4.3. An equilibrium x
∗
of x
/
= f (x) is called isolated if there exists a
positive number such that the open ball B(x
∗
, ) contains no equilibria other
than x
∗
.
Example 3.4.4. Find all equilibria of the system
x
/
1
= x
1
−2x
1
x
2
x
/
2
= x
1
x
2
−3x
2
.
Solution: Setting both x
/
1
= 0 and x
/
2
= 0, we seek points which simultaneously
satisfy
0 = x
1
−2x
1
x
2
= x
1
(1 −2x
2
)
0 = x
1
x
2
−3x
2
= x
2
(x
1
−3).
If we set x
1
= 0 in the ﬁrst equation, this would force x
2
= 0 in order for the
second equation to hold. Likewise, if we set x
2
= 1/2 in the ﬁrst equation,
nonli near systems: local theory 95
this forces x
1
= 3 in the second equation. It follows that there are exactly two
solutions of this system:
_
x
1
x
2
_
=
_
0
0
_
and
_
x
1
x
2
_
=
_
3
1
2
_
.
These are the equilibrium solutions, and both of them are isolated.
Remark. (i) Finding equilibria of nonlinear systems is usually impossible to
do algebraically. There is no straightforward extension of the matrix algebra
techniques you learned for handling linear, constantcoefﬁcient systems. (ii)
Whereas matrix systems Ax = b have either 0, 1, or inﬁnitely many solutions,
nonlinear systems can have any number of solutions. In the above example, there
were two equilibria (which could never happen for the linear systems discussed
in the previous chapter).
As we might expect, if we start out in equilibrium of a system x
/
= f (x),
then we are stuck there forever. More exactly, suppose the initial condition is
x(0) = x
∗
, where x
∗
is an equilibrium. Clearly the constant function x(t) = x
∗
is
a solution of the initial value problem, because x
/
= 0 and f (x) = 0. Assuming
that all partial derivatives in J f (x) are continuous in the vicinity of x
∗
, then the
Fundamental Existence and Uniqueness Theorem 3.2.2 guarantees that the initial
value problem has a unique solution in some interval containing t = 0. Since we
have already produced the solution x(t) = x
∗
, this constant solution must be the
only solution.
Linearization. Our earlier remarks about linear approximation of functions
allude to how we will approximate nonlinear systems of odes with linear ones.
For the sake of illustration, we introduce the process of linearization for two
variable systems
x
/
1
= f
1
(x
1
, x
2
), x
/
2
= f
2
(x
1
, x
2
), (3.5)
and remark that the following is easily extended to nvariable systems. Suppose
that (x
∗
1
, x
∗
2
) is any point in R
2
and we wish to approximate the behavior of the
nonlinear system (3.5) near this point. Replacing the righthand sides of the
96 equi li bri a and li neari zati on
equations (3.5) with their tangent plane approximations at (x
∗
1
, x
∗
2
), we obtain
_
x
/
1
x
/
2
_
≈
_
f
1
(x
∗
1
, x
∗
2
)
f
2
(x
∗
1
, x
∗
2
)
_
. ¸¸ .
vector in R
2
+
_
∇f
1
(x
∗
1
, x
∗
2
)
∇f
2
(x
∗
1
, x
∗
2
)
_
. ¸¸ .
22 Jacobian
_
x
1
−x
∗
1
x
2
−x
∗
2
_
. ¸¸ .
vector in R
2
.
This is nothing more than a restatement of Equation (3.3) in n = 2 dimensions.
If we use vector notation
x =
_
x
1
x
2
_
and x
∗
=
_
x
∗
1
x
∗
2
_
,
then the linear approximation of the system (3.5) becomes
x
/
≈ f (x
∗
) + J f (x
∗
)(x −x
∗
).
We give this approximation a name.
Deﬁnition 3.4.5. The system
x
/
= f (x
∗
) + J f (x
∗
)(x −x
∗
) (3.6)
is called the linearization of the system x
/
= f (x) at the point x = x
∗
.
The linearization is a linear, constantcoefﬁcient system which can be solved
using the techniques in the previous chapter. If x
∗
happens to be an equilib
rium of the system, then f (x
∗
) = 0 and the linearization takes the particularly
convenient form
x
/
= J f (x
∗
)(x −x
∗
).
Example 3.4.6. Consider our earlier example
x
/
1
= x
1
−2x
1
x
2
, x
/
2
= x
1
x
2
−3x
2
,
for which we found two equilibria:
_
0
0
_
and
_
3
1
2
_
.
nonli near systems: local theory 97
To obtain the linearization at the origin, we begin by computing the Jacobian
J f (x) =
_
1 −2x
2
−2x
1
x
2
x
1
−3
_
.
Evaluating the Jacobian at the ﬁrst equilibrium gives
J f (0, 0) =
_
1 0
0 −3
_
,
and therefore the linearization or our system at (0, 0) is
_
x
/
1
x
/
2
_
=
_
1 0
0 −3
_ _
x
1
x
2
_
.
We immediately see that the origin is a saddle for the linearized system and the
solution is x
1
(t) = c
1
e
t
and x
2
(t) = c
2
e
−3t
. The linearization about the other
equilibrium point is handled in a similar way. Referring to the Jacobian matrix
J f (x) above, we calculate that
J f
_
3,
1
2
_
=
_
0 −6
1
2
0
_
.
Notice that the characteristic equation of this matrix is λ
2
+ 3 = 0, which
means the eigenvalues are λ = ±
√
3i. This would suggest that the equilibrium
x
∗
= (3, 1/2) is a center, at least for the linearized system
_
x
/
1
x
/
2
_
=
_
0 −6
1
2
0
_ _
x
1
−3
x
2
−
1
2
_
=
_
0 −6
1
2
0
_ _
x
1
x
2
_
+
_
3
−
3
2
_
.
Solving this (barely inhomogeneous) system is straightforward.
When linearizing about an equilibrium, one would hope that the linearized
system would mimic the behavior of the nonlinear system, at least in the vicinity
of the equilibrium. This is usually, but not always, the case. In the previous
example, one of the equilibria was hyperbolic (no eigenvalues of the Jacobian
had zero real part), and the other was nonhyperbolic. In the next section, we
shall formally state the HartmanGrobman Theorem, which tells us that if we
linearize about a hyperbolic equilibrium, then the linearization exhibits the same
98 the hartman grobman theorem
qualitative behavior as the original nonlinear system. This need not be true for
nonhyperbolic equilibria.
3.5. The HartmanGrobman Theorem
In the previous section, we introduced the notion of isolated equilibria for
systems of odes.
Example 3.5.1. For a linear, homogeneous, constantcoefﬁcient system x
/
= Ax,
equilibria x
∗
must satisfy Ax
∗
= 0. If A is invertible, then the only equilibrium is
the origin, x
∗
= 0, which is clearly isolated. If A is not invertible, then solutions
of Ax
∗
= 0 form a subspace of R
n
of dimension at least 1 (i.e., the nullspace of
A is nontrivial). This implies that there are inﬁnitely many equilibria, none of
which are isolated.
In this section, we will assume that all equilibria are isolated. We also need
one more deﬁnition concerning equilibria of nonlinear systems.
Deﬁnition 3.5.2. An equilibrium x
∗
of the system x
/
= f (x) is called hyperbolic
if all eigenvalues of the Jacobian J f (x
∗
) have nonzero real part.
Example 3.5.3. By algebra, you can show that the system
x
/
= x −xy y
/
= −y + xy
has exactly two equilibria: (x, y) = (0, 0) and (x, y) = (1, 1). The Jacobian matrix
is
J f (x, y) =
_
1 −y −x
y −1 + x
_
,
from which we calculate that
J f (0, 0) =
_
1 0
0 −1
_
and J f (1, 1) =
_
0 −1
1 0
_
.
Since the eigenvalues of J f (0, 0) are 1 and −1, the equilibrium (0, 0) is hyperbolic.
However, the eigenvalues of J f (1, 1) are λ = ±i, both of which have zero real
part. Thus, (1, 1) is a nonhyperbolic equilibrium.
We now state one of our most important theorems about the qualitative
behavior of solutions of nonlinear systems of odes.
nonli near systems: local theory 99
Theorem 3.5.4. (HartmanGrobman). Suppose x
∗
is an isolated equilibrium
of a nonlinear system x
/
= f (x). Then in the vicinity of x
∗
, the linearization
x
/
= J f (x
∗
)(x −x
∗
) about that equilibrium has the same qualitative behavior as
the original nonlinear system.
More precisely, there exists a positive number such that, within the open
ball B(x
∗
, ), the phase portraits of the nonlinear system and its linearization are
topologically equivalent. In other words, there is a continuous, onetoone corre
spondence between phase portrait trajectories which preserves the orientation of
corresponding trajectories. We illustrate this via an example.
Example 3.5.5. Although the system
x
/
= −x y
/
= x
2
+ y
is nonlinear, it can be solved exactly, one equation at a time. Solving the ﬁrst
equation yields
x(t) = x
0
e
−t
,
where x
0
= x(0). Substituting this into the second equation and using the
variation of parameters formula, we ﬁnd that
y(t) = y
0
e
t
+
1
3
x
2
0
_
e
t
−e
−2t
_
.
With the assistance of a computer, we could sketch the phase portrait for the
nonlinear system by choosing various initial conditions (x
0
, y
0
) and plotting the
parametrized curve (x(t), y(t)) given by the above formulas. The phase portrait
is sketched in the left panel of Figure 3.2. Let us compare the dynamics of the
nonlinear system with that of its linearization. Clearly the only equilibrium of
this system is (x, y) = (0, 0). We linearize by computing the Jacobian
J f (x, y) =
_
−1 0
2x 1
_
,
from which
J f (0, 0) =
_
−1 0
0 1
_
.
100 the stable mani fold theorem
x
y
nonlinear
system
linearized
system
x
y
Figure 3.2. Comparing the phase portrait of a nonlinear system (left
panel) with that of its linearization (right panel).
The eigenvalues of J f (0, 0) are −1 and 1, neither of which has zero real part.
Hence, the origin is an isolated, hyperbolic equilibrium point. Moreover, since
the eigenvalues are real and of opposite sign, the origin is a saddle. The solution
of the linearized system
_
x
/
y
/
_
=
_
−1 0
0 1
_ _
x
y
_
is x = x
0
e
−t
and y = y
0
e
t
. You may check that the stable subspace E
s
is the xaxis
and the unstable subspace E
u
is the yaxis. The phase portrait for this saddle is
shown in the right panel of Figure 3.2. Comparing the two panels in Figure 3.2
serves to illustrate the HartmanGrobman Theorem. The phase portrait for the
nonlinear system is a “warped” version of the phase portrait of its linearization.
However, if we “zoom in” very close the origin, the two phase portraits would be
indistinguishable. Notice that the unstable subspace E
u
for the linearization also
happens to form a separatrix for the nonlinear system. However, the nonlinear
system “bends” the stable subspace E
s
into a parabola—this parabolic trajectory
does not form a subspace of R
2
. In the next section, we explain that the nonlinear
counterparts of stable, unstable, and center subspaces are called stable, unstable,
and center manifolds.
3.6. The Stable Manifold Theorem
A manifold is a generalization of the ﬂat objects (lines, planes, and hyperplanes)
that you learned about in linear algebra. The precise deﬁnition of a manifold
nonli near systems: local theory 101
is somewhat technical and is not required for our purposes; instead we give
some examples. A smooth curve such as a circle in the plane is an example of a
onedimensional manifold. If we “zoom in” on any point within the circle, what
we see would be indistinguishable from a small interval on the (onedimensional)
real line. Likewise, the surfaces in R
3
you learned about in multivariable calculus
would be examples of twodimensional manifolds. If we suitably magnify any
tiny region on the graph of the paraboloid f (x, y) = x
2
+ y
2
, then the result
would be virtually indistinguishable from a region within the “ﬂat” space R
2
.
We shall deal only with differentiable manifolds (imagine a surface which does
not have any sharp corners or ridges).
Below, we shall ﬁnd that near hyperbolic equilibria, we expect nonlinear
systems to have stable and unstable manifolds which have the same dimensions
and invariance properties of E
s
and E
u
for the linearized systems. First, we
review the concept of the ﬂow.
Suppose that x
/
= f (x) and assume that all partial derivatives in J f (x) are
continuous for all x ∈ R
n
. Given an initial condition x(0) = x
0
, Theorem 3.2.2
tells us that the initial value problem has a unique solution which exists in some
open interval containing t = 0. Let φ
t
(x
0
) denote the solution of this initial value
problem, deﬁned on the largest time interval containing t = 0 over which the
solution exists.
Deﬁnition 3.6.1. The set of all such functions φ
t
(x
0
) (i.e., for all possible choices
of x
0
) is called the ﬂow of the ode x
/
= f (x).
Notice that for each choice of x
0
, the function φ
t
(x
0
) deﬁnes a parametrized
curve in R
n
(parametrized by t), crossing the point x
0
at time t = 0. For the
linear systems we discussed in the previous chapter, we know that the phase
portraits may contain saddles, foci, nodes, and centers. The ﬂows for nonlinear
systems may exhibit much more interesting behavior.
Deﬁnition 3.6.2. A subset E ⊂ R
n
is called invariant with respect to the ﬂow if
for each x
0
∈ E, the curve φ
t
(x
0
) remains inside E for all time t over which the
solution actually exists.
For linear, homogeneous, constantcoefﬁcient systems x
/
= Ax, the subspaces
E
s
E
c
, and E
u
associated with the equilibrium at the origin are all invariant with
respect to the ﬂow. If we start from an initial condition inside any one of these
102 the stable mani fold theorem
subspaces, our solution trajectory remains conﬁned to that subspace for all time
t.
We now state another major theorem which, together with the Hartman
Grobman Theorem, provides much of our basis for understanding the qualitative
behavior of solutions of nonlinear odes.
Theorem 3.6.3. (Stable Manifold Theorem). Suppose that x
/
= f (x) is a system
of odes for which the Jacobian matrix J f (x) consists of continuous functions.
Further suppose that this system has an isolated, hyperbolic equilibrium point at
the origin, and that the Jacobian J f (0) has k eigenvalues with negative real part
and (n −k) eigenvalues with positive real part. Then
There is a kdimensional differentiable manifold W
s
which is (i) tangent to
the stable subspace E
s
of the linearized system x
/
= J f (0)x at the origin; (ii) is
invariant with respect to the ﬂow; and (iii) for all initial conditions x
0
∈ W
s
, we
have
lim
t→∞
φ
t
(x
0
) = 0.
There is an (n −k)dimensional differentiable manifold W
u
which is (i) tan
gent to the unstable subspace E
u
of the linearized system x
/
= J f (0)x at the
origin; (ii) is invariant with respect to the ﬂow; and (iii) for all initial conditions
x
0
∈ W
u
, we have
lim
t→−∞
φ
t
(x
0
) = 0.
Here, W
s
and W
u
are called the stable and unstable manifolds, respectively.
Remark. Although the statement of this Theorem is a bit wordy, it is not difﬁcult
to understand what is going on. Basically, it says that near a hyperbolic, isolated
equilibrium point, nonlinear systems produce objects W
s
and W
u
which are
“curvy versions” of the subspaces E
s
and E
u
for the linearized system. The
manifold W
s
is tangent to the subspace E
s
at the equilibrium. In the previous
example, W
s
was a parabola which was tangent to E
s
(the xaxis) at the origin
(see Figure 3.2). If we start from an initial condition inside W
s
, we will stay
inside W
s
for all time t, and we will approach the equilibrium as t →∞. Similar
statements hold for the unstable manifold.
Example 3.6.4. In the example we did in the previous section
x
/
= −x, y
/
= x
2
+ y,
nonli near systems: local theory 103
we noted that this nonlinear system has an exact solution
x(t) = x
0
e
−t
, y(t) =
_
y
0
+
1
3
x
2
0
_
e
t
−
1
3
x
2
0
e
−2t
.
The origin is the only equilibrium solution, and we already established that it
is a saddle. To determine W
s
, we need to determine which special choices of
initial conditions (x
0
, y
0
) would cause us to approach the origin as t →∞. In the
equation for x(t), we ﬁnd that
lim
t→∞
x
0
e
−t
= 0
regardless of our choice for x
0
. This imposes no restrictions on our choice of
initial conditions. Taking the same limit in the equation for y(t) is much more
interesting: in order to guarantee that
lim
t→∞
__
y
0
+
1
3
x
2
0
_
e
t
−
1
3
x
2
0
e
−2t
_
= 0
we must insist that
y
0
+
1
3
x
2
0
= 0,
because the exponential function e
t
will increase without bound as t →∞. No
other restrictions are necessary, because e
−2t
→ 0 as t → ∞. Therefore, we
conclude that the onedimensional stable manifold is given by the parabola
y = −x
2
/3. A similar argument shows that W
u
is the yaxis. Notice that W
s
is
tangent to the stable subspace E
s
(the xaxis) at the equilibrium point.
Example 3.6.5. Consider the nonlinear system
x
/
= −x +3y
2
, y
/
= −y, z
/
= 3y
2
+ z.
Solve this system and compute the stable and unstable manifolds for the equilib
rium at the origin.
Solution: This is very similar to the previous example. By computing the
linearization, you should convince yourself that the origin really is a hyperbolic
equilibrium. We see immediately that y(t) = y
0
e
−t
. Substituting this expression
104 the stable mani fold theorem
into the ﬁrst equation, we obtain a linear, inhomogeneous equation
x
/
+ x = 3y
2
0
e
−2t
.
The variation of parameters technique applies. Using
e
_
1 dt
= e
t
as an integrating factor, we ﬁnd that
e
t
_
x
/
+ x
_
= 3y
2
0
e
−t
,
or equivalently,
d
dt
_
e
t
x
_
= 3y
2
0
e
−t
.
Integrating both sides,
e
t
x = e
t
x
¸
¸
t=0
+
_
t
0
3y
2
0
e
−s
ds = x
0
+3y
2
0
_
−e
−s
_¸
¸
t
0
= x
0
+3y
2
0
(1 −e
−t
).
Multiplying both sides by e
−t
,
x(t) = x
0
e
−t
+ 3y
2
0
_
e
−t
−e
−2t
_
.
The equation for z is solved in a similar fashion, and the general solution of the
system is
x =
_
x
0
+3y
2
0
_
e
−t
− 3y
2
0
e
−2t
y = y
0
e
−t
z =
_
z
0
+ y
2
0
_
e
t
− y
2
0
e
−2t
.
The stable manifold W
s
consists of all initial conditions (x
0
, y
0
, z
0
) such that the
ﬂow guides us to the origin as t →∞. That is,
lim
t→∞
φ
t
(x
0
, y
0
, z
0
) = (0, 0, 0).
Notice that all of the exponential functions in the solution are decaying except
for the e
t
term in the z(t) equation. In order to guarantee that we approach the
nonli near systems: local theory 105
W
s
W
u
x
z
y
Figure 3.3. Sketch of the twodimensional stable manifold z = −y
2
and
the onedimensional unstable manifold (zaxis) from the example in the
text.
origin as t →∞, we need to coefﬁcient of e
t
to be 0. This forces
z
0
+ y
2
0
= 0.
If we graph z = −y
2
in R
3
, the result is a twodimensional manifold: a parabolic
sheet (see Figure 3.3). Similarly, solutions in the unstable manifold must approach
the origin as t → −∞. In this limit, we have e
t
→ 0 but e
−t
→∞ and e
−2t
→∞.
Requiring the coefﬁcients of both e
−t
and e
−2t
to be 0, it must be the case that
x
0
= 0 and y
0
= 0, while z
0
remains free. Therefore, the unstable manifold W
u
consists of all points on the zaxis.
We remark that when calculating W
s
and W
u
, it was important to express
the solution (x, y, z) of the odes in terms of the initial conditions (x
0
, y
0
, z
0
) as
opposed to introducing purely arbitrary constants C
1
, C
2
, and C
3
.
Calculating W
s
and W
u
by hand is usually impossible. There are methods for
obtaining successive approximations of these manifolds, but such techniques can
be very tedious.
3.7. NonHyperbolic Equilibria and Lyapunov Functions
The HartmanGrobman and Stable Manifold Theorems tell us that near an
isolated, hyperbolic equilibrium point x
0
, the behavior of the nonlinear system
106 non hyperboli c equi li bri a and lyapunov functi ons
x
/
= f (x) is qualitatively similar to that of its linearization x
/
= J f (x
0
) · (x −x
0
).
What about nonhyperbolic equilibria?
Example 3.7.1. We claim that the system
x
/
= −y −x(x
2
+ y
2
) y
/
= x −y(x
2
+ y
2
)
has exactly one equilibrium (the origin). To verify this, we set both x
/
= 0 and
y
/
= 0 to obtain
y = −x(x
2
+ y
2
) and x = y(x
2
+ y
2
).
Combining these two equations, we have
x = −x(x
2
+ y
2
)
2
.
Clearly (x
2
+ y
2
) ≥ 0 regardless of x and y, with equality only if x = y = 0. If it
were the case that x ,= 0, we could divide both sides of the latter equation by x
to obtain 1 = −(x
2
+ y
2
)
2
< 0, which is absurd. Therefore, it must be the case
that x = 0, from which it follows that y = 0 as well.
Next, we claim that the equilibrium (0, 0) is nonhyperbolic. The Jacobian
matrix associated with our system is
J f (x, y) =
_
−3x
2
−y
2
−1 −2xy
1 −2xy −x
2
−3y
2
_
,
and evaluating this matrix at the equilibrium point yields
J f (x, y) =
_
0 −1
1 0
_
.
The linearization of our original system is
_
x
/
y
/
_
=
_
0 −1
1 0
_ _
x
y
_
,
and the eigenvalues of the coefﬁcient matrix are λ = ±i. Both eigenvalues have
zero real part, implying that the equilibrium is nonhyperbolic and that the
nonli near systems: local theory 107
origin is a center for the linearized system (with circular trajectories in the phase
plane). Is this an accurate representation of the behavior of the nonlinear system?
Sneaky observation: The original system is easy to analyze if we introduce a
“radial” variable u(t) = x(t)
2
+ y(t)
2
. Notice that u measures the square of the
Euclidean distance from the origin for points on our solution curve (x(t), y(t)).
Taking the derivative,
du
dt
= 2x
dx
dt
+ 2y
dy
dt
= 2x[−y −x(x
2
+ y
2
)] + 2y[x −y(x
2
+ y
2
)]
= −2x
2
(x
2
+ y
2
) −2y
2
(x
2
+ y
2
) = −2(x
2
+ y
2
)(x
2
+ y
2
) = −2u
2
.
The equation u
/
= −2u
2
is easy to solve by separation of variables, and we ﬁnd
that
u(t) =
1
2t + C
,
where C is an arbitrary constant which we could solve for if given an initial
condition u
0
. Since u is, by deﬁnition, a nonnegative quantity, we know that
u
0
≥ 0. In fact, we may assume that u
0
> 0 because otherwise we would have
x
0
= y
0
= 0, placing us at the equilibrium for all time t. Setting t = 0 in the
equation for u(t), we ﬁnd that C = u
−1
0
> 0 because u
0
> 0. Since C is positive,
the solution u(t) exists for all positive t because the denominator 2t + C is never
0. Taking the limit t →∞, we conclude that
lim
t→∞
u(t) = 0
regardless of our choice of initial conditions. This means that all solutions trajectories
for the nonlinear system approach the equilibrium at the origin as t → ∞.
Consequently, the origin is an attractor in spite of the fact that the linearized
system has a center at the origin. Figure 3.4 shows a sample trajectory in the
phase plane for the nonlinear system. The phase portrait for the linearized
system is exactly as in Figure 2.8e.
Deﬁnition 3.7.2. In the above example, the origin is called a weak sink or weak
attractor.
The reason for using the word “weak” in this deﬁnition is because trajectories
approach the origin slower than those of “normal” attractors, which approach
exponentially fast.
108 non hyperboli c equi li bri a and lyapunov functi ons
x
y
Figure 3.4. Phase plane of a nonlinear system for which the origin is a
weak attractor (see text).
Example 3.7.3. Changing signs in the previous example, the system
x
/
= −y + x(x
2
+ y
2
) y
/
= x + y(x
2
+ y
2
)
still has a nonhyperbolic equilibrium at the origin. This time, the origin is
unstable (a repeller), whereas the origin is a [stable] center for the linearized
system. Deﬁning u(t) as before, the reader is encouraged to show that u(t)
2
actually blows up to ∞ in a ﬁnite amount of time for any choice of initial
conditions (excluding the equilibrium itself). This is a major contrast with the
behavior of solutions of the linearized system, which exist and remain bounded
for all real t.
Moral: The two examples above should convince you that near nonhyperbolic
equilibria, we need a completely different approach for classifying stability.
Earlier, we gave an intuitive, but somewhat vague deﬁnition of what it means for
an equilibrium to be stable. We now give a precise mathematical deﬁnition for
stability of equilibria. To make the deﬁnition slightly less technical to state, we
will assume that the ﬂow φ
t
is deﬁned for all positive t. You may wish to consult
Deﬁnition 3.4.2 for a reminder of how the open ball B(x
∗
, ) is deﬁned.
Deﬁnition 3.7.4. Suppose x
∗
is an isolated equilibrium of the ode x
/
= f (x) and
let x
0
denote an initial condition. Then x
∗
is called
locally stable if given any number > 0 there exists a δ > 0 such that whenever
x
0
∈ B(x
∗
, δ), then the solution satisﬁes x(t) ∈ B(x
∗
, ) for all t > 0.
nonli near systems: local theory 109
x
o
δ
ε
x*
Figure 3.5. Illustration of local stability of an equilibrium. The inner and
outer dashed circles have radii δ and , respectively, and are centered at
an equilibrium x
∗
.
locally asymptotically stable if x
∗
is locally stable and there exists η > 0 such
that
lim
t→∞
x(t) = x
∗
whenever x
0
∈ B(x
∗
, η).
unstable if x
∗
is not locally stable.
The notion of local stability is illustrated in Figure 3.5, and in words, this
concept is explained as follows. Suppose we are given a positive number . Then
no matter how small is, we can always produce another positive number δ such
that whenever we start from an initial condition whose distance from x
∗
is less
than δ, the solution curve x(t) will always have distance less than from x
∗
. In
some sense, this basically means that trajectories that start reasonably close to x
∗
can never “escape too far” from x
∗
. Local asymptotic stability means that if the
initial condition x
0
is appropriately close to x
∗
, then the solution curve x(t) will
actually converge to x
∗
as t → ∞. For linear systems, centers are examples of
equilibria that are stable but not asymptotically stable. The HartmanGrobman
Theorem and Stable Manifold Theorems ensure that any hyperbolic equilibrium
point of x
/
= f (x) is either locally asymptotically stable (if all eigenvalues of the
Jacobian matrix J f (x
∗
) have negative real part) or unstable (if any eigenvalue has
positive real part). An asymptotically stable equilibrium is sometimes called a
sink or attractor.
110 non hyperboli c equi li bri a and lyapunov functi ons
Lyapunov Functions. Classifying nonhyperbolic equilibria x
∗
as stable, asymp
totically stable, or unstable can be incredibly difﬁcult (and often impossible).
We now describe a classiﬁcation technique that was originally proposed by
Russian mathematician A.M. Lyapunov in his 1892 doctoral dissertation. Con
sider the ode x
/
= f (x), where f is continuously differentiable. A solution x(t)
of this equation can be written in terms of its component functions, x(t) =
[x
1
(t), x
2
(t), . . . x
n
(t)]. Now suppose that V : R
n
→R is a continuously differen
tiable scalarvalued function. Then by the chain rule,
d
dt
V(x(t)) =
d
dt
V(x
1
(t), x
2
(t), . . . x
n
(t))
=
∂V
∂x
1
dx
1
dt
+
∂V
∂x
2
dx
2
dt
+ · · · +
∂V
∂x
n
dx
n
dt
=
_
∂V
∂x
1
,
∂V
∂x
2
, . . .
∂V
∂x
n
_
•
_
dx
1
dt
,
dx
2
dt
, . . . ,
dx
n
dt
_
(a dot product)
= ∇V(x) • x
/
(t) = ∇V(x) • f (x).
Observation. This calculation tells us how the function V changes as we move
along a solution curve x(t). In particular, if we ﬁnd that
d
dt
V(x(t)) < 0
inside some set E ⊂ R
n
, then the function V decreases as we move along solution
curves in E in the direction of increasing t.
Lyapunov exploited this observation to provide a creative but intuitive way
for analyzing stability of equilibria x
∗
. The idea is to deﬁne a function V on a
set E containing x
∗
, where V is chosen in such a way that we can tell whether
the ﬂow in E is towards or away from the equilibrium. In what follows, we have
in mind a system x
/
= f (x) with an isolated equilibrium x
∗
. We assume that
f is continuously differentiable in some open ball E = B(x
∗
, ) of radius > 0
centered at the equilibrium.
Theorem 3.7.5 (Lyapunov). Suppose there exists a function V : R
n
→R which is
(i) deﬁned and continuously differentiable on the set E = B(x
∗
, ); (ii) V(x
∗
) = 0;
and (iii) V(x) > 0 if x ,= x
∗
. Then the equilibrium x
∗
is
Stable if
d
dt
V(x(t)) = ∇V(x) • f (x) ≤ 0
nonli near systems: local theory 111
for all x ∈ E.
Asymptotically stable if
d
dt
V(x(t)) = ∇V(x) • f (x) < 0
for all x ∈ E, except possibly at x
∗
itself.
Unstable if
d
dt
V(x(t)) = ∇V(x) • f (x) > 0
for all x ∈ E, except possibly at x
∗
itself.
Deﬁnition 3.7.6. Any function V : R
n
→ R satisfying the conditions of Theo
rem 3.7.5 is called a Lyapunov function.
Some geometric considerations may help convince you that Lyapunov’s Theo
rem is intuitively plausible. Figure 3.6 illustrates the shape of a typical Lyapunov
function for a twovariable system of odes with an equilibrium at the origin.
In this case, the graph of the Lyapunov function V(x, y) is a surface in R
3
, and
is positive everywhere except at the equilibrium (where it is 0). Any solution
(x(t), y(t)) of the odes deﬁnes a parametrized curve in the xyplane, and there
fore V(x(t), y(t)) deﬁnes a curve on the surface V. If V(x(t), y(t)) decreases to
0 as t increases, then the corresponding solution trajectory in the xyplane is
“funneled” towards the origin, and we conclude that the equilibrium must be
asymptotically stable. In general, ﬁnding a Lyapunov function is very difﬁcult,
and there is no general procedure for doing so.
Example 3.7.7. In a previous example, we showed that the planar system
x
/
= −y −x(x
2
+ y
2
) y
/
= x −y(x
2
+ y
2
)
has exactly one equilibrium (the origin), and its linearization about the origin is
_
x
/
y
/
_
=
_
0 −1
1 0
_ _
x
y
_
.
Since the eigenvalues of the coefﬁcient matrix are λ = ±i, both of which have
zero real part, the origin is a nonhyperbolic equilibrium. We also proved that the
origin is actually asymptotically stable, by introducing a new variable u = x
2
+y
2
and solving a differential equation for u. We now give an alternative proof using
112 non hyperboli c equi li bri a and lyapunov functi ons
start
start
V(x,y)
xy plane
Figure 3.6. Illustration of a Lyapunov function for a twovariable system
with an equilibrium at the origin. Corresponding to each solution of the
ode in the phase plane, there is a curve on the surface V(x, y). If all such
curves on the surface V decrease to 0 as t →∞, then the corresponding
solution trajectories in the xyplane must all approach the equilibrium
as t →∞, implying that the equilibrium is asymptotically stable.
Theorem 3.7.5. We claim that V(x, y) = x
2
+ y
2
is a Lyapunov function for the
equilibrium at the origin. Clearly V(0, 0) = 0, and V(x, y) > 0 if (x, y) ,= (0, 0).
Moreover, note that ∇V(x, y) • f (x, y) is given by
(2x, 2y) • (−y −x(x
2
+ y
2
), x −y(x
2
+ y
2
))
= −2xy −2x
2
(x
2
+ y
2
) +2xy −2y
2
(x
2
+ y
2
) = −2(x
2
+ y
2
)
2
.
Since ∇V(x, y) • f (x, y) < 0 except at (0, 0), we conclude from Lyapunov’s
Theorem that the origin is asymptotically stable.
One interesting observation regarding the previous example is that the Lya
punov function V(x, y) is deﬁned for all points (x, y) ∈ R
2
, and that the inequal
ity ∇V(x, y) • f (x, y) < 0 is satisﬁed in the entire plane (except at the origin).
This means that starting from any initial conditions, solution trajectories always
approach the origin as t → ∞. In other words, we may actually conclude that
the origin is globally asymptotically stable, not just locally stable as the Theorem
would guarantee. Among other things, this tells us that the above system of odes
cannot possess any periodic solutions.
For certain types of systems of odes, Lyapunov functions are readily available.
We now discuss one such example.
nonli near systems: local theory 113
Gradient systems. Gradient systems are very special types of systems of odes
which have the form
dx
k
dt
= −
∂V
∂x
k
k = 1, 2, . . . , n,
where V : R
n
→ R is a continuously differentiable function. In other words,
the right hand sides of the odes in the system are obtained by taking partial
derivatives of the same underlying function V. Now suppose that x
/
= f (x) is
a gradient system with an isolated equilibrium x
∗
, and observe that
d
dt
V(x(t)) = ∇V(x) • f (x)
=
_
∂V
∂x
1
,
∂V
∂x
2
, . . . ,
∂V
∂x
n
_
•
_
−
∂V
∂x
1
, −
∂V
∂x
2
, . . . , −
∂V
∂x
n
_
= −
n
∑
k=1
_
∂V
∂x
k
_
2
= −∇V(x)
2
2
,
which could never be positive. Consequently, if V(x) ≥ 0 for all x ∈ R
n
with
equality only if x = x
∗
, then the function V itself would be a Lyapunov function
for the system. We could immediately conclude that the equilibrium x
∗
is stable.
Example 3.7.8. Consider the rather silly system x
/
= −4x
3
and y
/
= −4y
3
, which
can be easily solved by hand using separation of variables. The only equilibrium
solution is at the origin (x, y) = (0, 0). The Jacobian matrix associated with this
system is
J f (x, y) =
_
−12x
2
0
0 −12y
2
_
,
and evaluating this matrix at the equilibrium point yields
J f (0, 0) =
_
0 0
0 0
_
.
The linearization is completely useless, because every point in R
2
is an equilib
rium solution of
_
x
/
y
/
_
=
_
0 0
0 0
_ _
x
y
_
.
Notice that λ = 0 is a double eigenvalue of the matrix J f (0, 0), and the equilib
rium is nonhyperbolic. We claim that the original nonlinear system is a gradient
114 non hyperboli c equi li bri a and lyapunov functi ons
system. Indeed, suppose that V(x, y) is a function with the property that
−
∂V
∂x
= −4x
3
and −
∂V
∂y
= −4y
3
.
Integrating the ﬁrst of these equations with respect to x, we ﬁnd that
V(x, y) =
_
4x
3
dx = x
4
+ g(y),
where g is an arbitrary function of y. Imposing the requirement that ∂V/∂y = 4y
3
,
we calculate
∂
∂y
_
x
4
+ g(y)
_
= 4y
3
,
which means that g
/
(y) = 4y
3
. Integrating again, we have g(y) = y
4
+C, where
C is an arbitrary constant. In order for our function V(x, y) = x
4
+ y
4
+ C to be
a Lyapunov function, we must insist that C = 0 so that V(0, 0) = 0. Moreover,
∇V(x, y) • f (x, y) = (4x
3
, 4y
3
) • (−4x
3
, −4y
3
) = −16(x
6
+ y
6
) < 0
for all (x, y) ,= (0, 0). By Lyapunov’s Theorem, we conclude that the origin is
asymptotically stable. In fact, since the above calculations hold in the entire
plane R
2
, the origin is globally asymptotically stable.
Remark. Given a function f : R
2
→ R
2
, any function V : R
2
→ R with the
property that ∇V = f is called a potential function for f . It is not always possible
to ﬁnd a potential function for a given f . Most multivariable calculus textbooks
state conditions under which a potential function exists.
In your homework exercises, you will consider more “exotic” Lyapunov func
tions than the one used in the toy example above.
Exercises
1. Consider the function f : R
2
→R
2
deﬁned by
f (x, y) =
_
f
1
(x, y)
f
2
(x, y)
_
=
_
xy cos(ln y)
arctan
_
x
2
y
_
_
.
Find the linear approximation for f at the point (x
0
, y
0
) = (1, 1).
nonli near systems: local theory 115
2. Recall from an example that the initial value problem
dy
dx
= y
2
, y(0) = 1 has
solution y = (1 −x)
−1
. Although this solution is not wellbehaved, we do get
one minor consolation prize: Show that the solution is unique.
3. Show that the initial value problem
dx
dt
= arctan(xy) x(0) = x
0
,
dy
dt
= xe
−(x
2
+y
2
)
y(0) = y
0
has a unique solution regardless of the values of x
0
and y
0
.
4. Suppose that f (x) is continuous on some closed interval [a, b]. Show that
¸
¸
¸
¸
_
b
a
f (x) dx
¸
¸
¸
¸
≤
_
b
a
 f (x) dx.
5. This exercise will guide you through a proof of Gronwall’s Inequality, which
is stated as follows. Let f (t) be a nonnegative continuous function, and
suppose that there exist positive constants C and K such that
f (t) ≤ C + K
_
t
0
f (s) ds
for all t ∈ [0, a]. Then f (t) ≤ Ce
Kt
for all t ∈ [0, a].
(a) Deﬁne the function F(t) = C + K
_
t
0
f (s) ds for t ∈ [0, a]. Explain why
F(t) ≥ f (t) and why F(t) > 0 for all t ∈ [0, a].
(b) The Fundamental Theorem of Calculus shows that F
/
(t) = Kf (t). Com
bining this with part (a) above, show that, for all t ∈ [0, a],
F
/
(t)
F(t)
≤ K.
(c) Making the sneaky observation that
F
/
(t)
F(t)
=
d
dt
ln(F(t)),
use the result of part (b) above to show that
ln F(t) ≤ Kt +ln F(0)
116 non hyperboli c equi li bri a and lyapunov functi ons
for all t ∈ [0, a]. Finally, exponentiate both sides of this inequality and
recall from part (a) that f (t) ≤ F(t). Gronwall’s Inequality follows.
6. Find all equilibria of the system
dx
dt
= x −y
2
dy
dt
= x
2
y −4y.
Linearize the system at each hyperbolic equilibrium, and classify these equilib
ria as stable/unstable nodes, saddles, or stable/unstable foci.
7. Find all equilibria of the system
dx
dt
= (x
2
−1)(y −1)
dy
dt
= (x −2)y.
Linearize the system at each hyperbolic equilibrium and classify these equilib
ria. Try to draw the phase portrait for the system. To do so, start by using the
linearized systems to sketch trajectories in the vicinity of each hyperbolic equi
librium. Then, try to “interpolate”, by ﬁlling in other trajectories throughout
the phase plane. If you have access to computer software which can generate
the true phase portrait, feel free to see how it compares to your hand sketch.
8. Consider the system
dx
dt
= −y + x(µ −x
2
−y
2
)
dy
dt
= x + y(µ −x
2
−y
2
),
where µ is a constant.
(a) Show that the origin is the only equilibrium of this system.
(b) Find the linearization of this system at the origin.
(c) Note that the value of µ determines how the phase plane of the linearized
system looks. What sort of equilibrium do we have if µ > 0? What if
µ < 0? What if µ = 0?
9. Sketch the phase planes for the following system and its linearization at the
origin.
x
/
= −y + x(x
2
+ y
2
) y
/
= x + y(x
2
+ y
2
)
When dealing with the nonlinear system, you may want to deﬁne u(t) =
x(t)
2
+ y(t)
2
and proceed as in an example in the text.
nonli near systems: local theory 117
10. Consider the system
x
/
= −3x y
/
= x
2
+2y,
which has an isolated equilibrium at the origin. Sketch the phase portraits for
this system and its linearization at the origin. Find equations for the stable
and unstable manifolds W
s
and W
u
.
11. The origin is the only equilibrium of the system
dx
dt
= 2x +5y
3
,
dy
dt
= −y.
Find the stable and unstable manifolds W
s
(0, 0) and W
u
(0, 0) and sketch the
phase plane.
12. Find all equilibria of the system
dx
dt
= −x
dy
dt
= 2(x
3
+ y)
For each equilibrium, ﬁnd formulas for the stable and unstable manifolds.
13. This problem concerns a massspring system (see ﬁgure). An object of mass
m is attached to a spring. The object slides back and forth on a table, eventually
returning to its resting position due to friction.
x
0
0
rest
stretched
Let x denote the displacement of the object from its equilibrium (x = 0)
position, and let v denote its velocity. We will use the following assumptions
when modeling the motion:
The force exerted on the object by the spring is proportional to the dis
placement x of the object from equilibrium. The spring’s force acts in the
opposite direction of the displacement. In other words, we have assumed that
118 non hyperboli c equi li bri a and lyapunov functi ons
Hooke’s Law holds: the force exerted by the spring is given by F
spring
= −kx
where k > 0 is a positive constant.
We assume that there is a frictional force exerted on the object as well, and
this force acts in the direction opposite the velocity vector. For small velocity v,
the friction is roughly proportional to the velocity, and for large v, the friction
is proportional to the square of the velocity. That is, F
frictional
= −Av − Bv
2
,
where A and B are positive constants.
According to Newton’s Law, F = ma, where F is the total force acting on
the object, m is mass, and a =
dv
dt
is acceleration. In other words,
ma = F
spring
+ F
frictional
= −kx − Av −Bv
2
.
For convenience, we will assume that mass has been scaled so that m = 1.
Since v =
dx
dt
and a =
dv
dt
, we get a system of two odes for displacement and
velocity:
dx
dt
= v
dv
dt
= −kx − Av −Bv
2
.
(a) Show that this system has only one equilibrium, and ﬁnd it. Give a
physical interpretation of the equilibrium.
(b) Find the linearization of the system about the equilibrium.
(c) Give an intuitive physical explanation for why you would expect this
equilibrium to be stable. Then, rigorously show that the equilibrium is
asymptotically stable (and therefore hyperbolic), regardless of the choices
of the positive constants A, B, and k.
(d) Solve the linearized system. If you do this correctly, you will ﬁnd that the
exact form of the solution depends upon the sign of A
2
−4k.
(e) Give a physical interpretation of the differences between the solutions
you would get if A
2
−4k is positive versus those you would get if A
2
−4k
is negative. The constant k is called the spring constant and its value
determines the stiffness of the spring.
14. The LotkaVolterra equations model the populations of two interacting
species. In what follows, x(t) denotes the population of a prey species, and
y(t) denotes the population of a predator species which depends upon the
nonli near systems: local theory 119
prey for survival. The equations of this model are:
dx
dt
= Ax −Bxy
dy
dt
= Cxy −Dy
where A, B, C, and D are all positive constants. As usual, denote the right
hand side of this system by f (x, y), which maps R
2
→R
2
.
(a) Note that the origin (x, y) = (0, 0) is an equilibrium. Find the linearization
about this equilibrium and solve the linearized system. Classify the
equilibrium as stable/unstable and determine whether it is a saddle, node,
focus, or center.
(b) There is one other equilibrium (x
∗
, y
∗
) of this system. Find it and show
that it is nonhyperbolic. Do not bother to solve the linearized system.
(c) Let V(x, y) = Cx −Dln x + By − Aln y + E, where E is a constant. Show
that ∇V · f = 0.
(d) If (x
∗
, y
∗
) is the equilibrium you found in part (b) above, how should you
choose the constant E so that V(x
∗
, y
∗
) = 0?
(e) Use the second derivative test to show that V(x, y) has a local minimum at
the equilibrium you found in part (b). Conclude that V(x, y) is a Lyapunov
function for this equilibrium and, using your result from part (c), state
what you have learned about the stability of this equilibrium.
15. Consider the system
dx
dt
= 4x(x −1)
_
x −
1
2
_
dy
dt
= 2y.
(a) Find all equilibria of this system.
(b) Show that V(x, y) = x
2
(x +1)
2
+y
2
is a Lyapunov function for this system
at the origin.
(c) Use V(x, y) to classify the stability of the origin.
16. Given that
dx
dt
= x
2
y −
xy
2
2
dy
dt
= x
2
_
x
3
−
y
2
_
is a gradient system, ﬁnd a Lyapunov function V(x, y) such that
dx
dt
= −∂V/∂x
and
dy
dt
= −∂V/∂y. Then, use V(x, y) to prove that the origin is asymptotically
stable.
120 non hyperboli c equi li bri a and lyapunov functi ons
17. Consider the nonlinear system
dx
dt
= y
dy
dt
= −a sin x,
where a is a positive constant.
(a) Show that the origin is a nonhyperbolic equilibrium.
(b) Deﬁne V(x, y) = a(1 −cos x) +
1
2
y
2
. Show that this is a Lyapunov function
for the system (at least in some appropriately chosen region of the xy
plane which contains the origin).
(c) What can you conclude about the stability of the origin?
18. Consider the nonlinear system
dx
dt
= −y + x(x
2
+ y
2
)
dy
dt
= x + y(x
2
+ y
2
),
which looks suspiciously like an example in the text (hint, hint). Notice that
the system has a unique equilibrium (the origin), and this equilibrium is
nonhyperbolic. Find a Lyapunov function and use it to determine whether
the origin is stable or not.
19. The nullclines of a nonlinear system of odes provide a useful way of under
standing the dynamics. For a planar system of the form
dx
dt
= f
1
(x, y)
dy
dt
= f
2
(x, y),
the xnullcline is the curve deﬁned by f
1
(x, y) = 0 and the ynullcline is the
curve deﬁned by f
2
(x, y) = 0. The intersections of these nullclines are the
equilibria of the system.
Consider the system
dx
dt
= −αx + y
dy
dt
=
5x
2
4 + x
2
−y,
where α is a positive parameter.
(a) Find and plot the nullclines for this system with α = 2. Since the origin
is the only intersection point of the x and ynullclines, it is the only
equilibrium. Linearize the system at origin and classify the equilibrium.
(b) Find and plot the nullclines for this system with α = 1. How many
equilibria does the system now have?
nonli near systems: local theory 121
(c) Again assuming that α = 1, ﬁnd the equilibria algebraically.
(d) What types of equilibria are these?
(e) Notice that changing α from 1 to 2 affects the number of equilibria. In fact,
there is some critical value of α (between 1 and 2) for which the system
experiences a signiﬁcant qualitative change (i.e., a change in the number
of equilibria). Find this critical value of α. This critical value is called
a bifurcation point. In Chapter 5, we will explore this dependence on
parameters in more detail.
CHAPTER 4
Periodic, Heteroclinic, and Homoclinic Orbits
I
n this chapter, we shift our attention away from equilibria, instead seeking
more “interesting” solutions of nonlinear systems x
/
= f (x). Much of our
discussion involves planar systems (i.e., f : R
2
→ R
2
), because such systems
admit particularly simple criteria for the existence of periodic solutions.
4.1. Periodic Orbits and the PoincaréBendixon Theorem
A nonequilibrium solution x of the system x
/
= f (x) is periodic if there exists a
positive constant p such that x(t + p) = x(t) for all time t. The least such p is
called the period of the solution, and tells us how often the solution trajectory
“repeats itself”. In the phase portrait for the system of odes, periodic solutions
(sometimes called periodic orbits) always appear as closed curves. On the other
hand, not every closed curve corresponds to a periodic solution, as we shall see
when we discuss homoclinic orbits.
For linear, constantcoefﬁcient systems, we learned to associate pure imaginary
eigenvalues with periodic solutions. Determining whether a nonlinear system
has periodic solutions is less straightforward. In preparation for stating criteria
for existence of periodic solutions, we review some basic notions from calculus.
Suppose that Γ(t) = (γ
1
(t), γ
2
(t)) is a parametrized curve in R
2
; i.e., x =
γ
1
(t) and y = γ
2
(t). Assume that γ
1
(t) and γ
2
(t) are continuously differentiable.
At a given time t = t
0
, the tangent vector to the curve Γ(t) is given by Γ
/
(t
0
) =
(γ
/
1
(t
0
), γ
/
2
(t
0
)).
122
peri odi c, heterocli ni c, and homocli ni c orbi ts 123
Γ(t)
(t )
o
Γ
Γ(t)
n
τ
inward
outward
(a) (b)
Figure 4.1. (a) Illustration of a tangent vector τ and normal vector n to
a curve Γ at a particular time t = t
0
. (b) For the closed curves that we
shall consider, each point on the curve has exactly one outward unit
normal vector and one inward unit normal vector.
Deﬁnition 4.1.1. Any nonzero vector n(t
0
) in R
2
which is perpendicular to the
tangent vector Γ
/
(t
0
) is called a normal vector to the curve at t = t
0
. If n
2
= 1,
then n is called a unit normal vector.
Figure 4.1a illustrates the normal and tangent vectors to a parametrized curve Γ.
When we deal with closed curves (Figure 4.1b), we will always presume that the
curves are sufﬁciently “wellbehaved” (smooth) that there is a unique inward
unit normal vector and unit outward normal vectors at each point along the
curve. The following Lemma formally states an observation that was made when
we discussed phase portraits in Chapter 2.
Lemma 4.1.2. Consider a systemx
/
= f (x) where f is continuously differentiable.
Then solution trajectories in the phase portrait cannot intersect each other.
Proof. Exercise. Convince yourself that if two trajectories did intersect, then this
would violate the Fundamental Existence and Uniqueness Theorem 3.2.2.
With the above Lemma in mind, we now give an instructive example that will
motivate our main result regarding existence of periodic solutions.
Example 4.1.3. The system
x
/
= f
1
(x, y) = −y + x(1 −x
2
−y
2
)
y
/
= f
2
(x, y) = x + y(1 −x
2
−y
2
)
124 peri odi c orbi ts and the poi ncaré bendi xon theorem
has exactly one equilibrium solution: (x, y) = (0, 0). The linearization about the
equilibrium is given by
_
x
/
y
/
_
=
_
1 −1
1 1
_ _
x
y
_
.
The coefﬁcient matrix is in real canonical form, and has eigenvalues λ = 1 ± i.
Since the origin is a hyperbolic equilibrium, the HartmanGrobman Theorem
tells us that our original nonlinear system has an unstable focus at the origin.
Observation. Even though the origin is a repeller, solution trajectories cannot
escape too far from the origin as t → ∞. Notice that x
2
+ y
2
measures the
square of the Euclidean distance from the origin. If x
2
+ y
2
becomes too large,
an inspection of the right hand sides of our odes suggests that we will be pulled
back towards the origin.
To make this claim more rigorous, we will construct a curve Γ which simulta
neously encloses the origin and has the property that solution trajectories in the
phase plane always cross Γ from outside to inside as t increases. Suppose that Γ
parameterizes a circle of radius R centered at the origin. If we agree to orient the
curve Γ counterclockwise, then one natural parametrization would be
Γ(t) = (γ
1
(t), γ
2
(t)) = (Rcos t, Rsin t).
At each time t, the unit outward normal vector is simply n(t) = (cos t, sin t), as
illustrated in Figure 4.2.
Question: If choose some point (x, y) = (Rcos t, Rsin t) on the curve Γ, does
the ﬂow of the system of odes direct our motion towards the interior or exterior
of the curve?
Answer: We claim that if f (x, y) • n(t) < 0, then the ﬂow is directed inward,
and if f (x, y) • n(t) > 0, then the ﬂow is directed outward. To see why, we recall
the geometric interpretation of the dot product of two vectors:
f (x, y) • n(t) = f (x, y)
2
n
2
cos θ,
peri odi c, heterocli ni c, and homocli ni c orbi ts 125
where θ is the angle between the two vectors f (x, y) and n. Clearly f (x, y)
2
and n
2
are positive since they represent lengths of vectors. If π/2 < θ < 3π/2,
then the vector f (x, y) must be oriented towards the interior of the curve Γ
(see Figure 4.2). For this range of angles θ, we have cos θ < 0 and therefore
f (x, y) • n(t) < 0 as well.
In the present example, our function f (x, y) is given by
_
f
1
(x, y)
f
2
(x, y)
_
=
_
−y + x(1 −x
2
−y
2
)
x + y(1 −x
2
−y
2
)
_
,
and n = (cos t, sin t). Computing f • n at (x, y) = (Rcos t, Rsin t),
f (x, y) • n
= [−Rsin t + R(cos t)(1 −R
2
), Rcos t + R(sin t)(1 −R
2
)] • (cos t, sin t),
where we have used the fact that x
2
+y
2
= R
2
. Expanding the dot product yields
f (x, y) • n = −Rsin t cos t + R(cos
2
t)(1 −R
2
) + Rcos t sin t + R(sin
2
t)(1 −R
2
)
= R(cos
2
t)(1 −R
2
) + R(sin
2
t)(1 −R
2
) = R(1 −R
2
).
Observe that if the radius R of our circular curve Γ is larger than 1, then
f • n < 0, implying that the ﬂow is directed inward. Likewise, if R < 1, then
f • n > 0, implying that the ﬂow is directed outward. In summary, we have
noted that
The only equilibrium of the system is (0, 0), and it is a repeller.
Excluding the equilibrium, all trajectories of this system are “funneled” to
wards a circle of radius 1 centered at the origin as t increases.
Different trajectories cannot intersect (Lemma 4.1.2).
It turns out that the only way to reconcile all three of these constraints is if the
circle of radius 1 centered at the origin corresponds to a very special type of a
solution: a periodic orbit.
In the preceding example, we constructed curves Γ in the plane R
2
and
then determined whether the ﬂow of the system of odes was directed towards
the interior or exterior of Γ as t increases. The notion that trajectories may
become “trapped” on the interior of some specially chosen curve Γ motivates
126 peri odi c orbi ts and the poi ncaré bendi xon theorem
f
n
θ
Figure 4.2. The outward unit normal vector for the circle (Rcos t, Rsin t)
is given by n = (cos t, sin t).
an important theorem. In what follows, we consider a planar system x
/
= f (x)
where x = (x
1
, x
2
) and f : R
2
→R
2
is continuously differentiable.
Theorem 4.1.4 (PoincaréBendixon.). Suppose that Γ is a continuously differen
tiable closed curve in R
2
and let n denote an outward unit normal vector at a
point (x
1
, x
2
) on Γ. If the ode has no equilibria on or inside Γ and f • n < 0 at
all points (x
1
, x
2
) on Γ, then there exists at least one periodic solution of the ode
inside Γ.
Remark. The PoincaréBendixon Theorem is valid only in two dimensions. In
three dimensions, the famous Lorenz odes have solutions which remain conﬁned
to a bounded set, but never approach an equilibrium or periodic orbit. In fact,
the Lorenz equations exhibit chaotic behavior.
If we can construct a curve Γ for which the conditions of the PoincaréBendixon
Theorem are satisﬁed, then any solution starting from an initial condition x
0
inside Γ must stay trapped inside Γ for all time t. As a consequence, we have
Corollary 4.1.5. If the conditions of the PoincaréBendixon Theorem are satisﬁed,
then any solution curve starting from an initial condition x
0
inside Γ will exist
for all time t ≥ 0.
Proof. In our planar system x
/
= f (x), we always assume that f : R
2
→ R
2
is
continuously differentiable. The function g : R
2
→R deﬁned by g(x) = x
2
is
also a continuous function, and therefore the composition g( f (x)) = f (x)
2
is
continuous. If Ω denotes the region enclosed by the continuous, closed curve Γ,
peri odi c, heterocli ni c, and homocli ni c orbi ts 127
then Ω is a closed and bounded set. It follows that the function f (x)
2
achieves
some maximum value B in the region Ω. Since f 
2
≤ B in this region, we may
apply Theorem 3.3.13 (with K = 0) to infer global existence for all t ≥ 0.
When attempting to prove that a planar system x
/
= f (x) has a periodic
solution, it can be very challenging to actually ﬁnd a curve Γ which satisﬁes
the conditions of the PoincaréBendixon Theorem. However, there is one piece
of good news—the theorem actually holds even if the curve Γ is only piecewise
differentiable. Consequently, instead of using smooth curves like circles, it is
possible to use polygonal paths (such as squares). You may also ﬁnd it helpful to
use computer software to sketch the vector ﬁeld deﬁned by the function f to see
whether it is possible to draw a closed curve Γ on which all arrows of the vector
ﬁeld are directed inward.
Like equilibria, periodic solutions of odes can be stable or unstable. A periodic
orbit p(t) is called asymptotically stable if all nearby solution trajectories approach
the periodic orbit as t → ∞. This deﬁnition could be made more rigorous, as
in our deﬁnition of stability of equilibria. However, we shall be content with
a more intuitive understanding of stability of periodic orbits. As an example,
consider an idealized frictionless pendulum which simply swings back and forth
without any external forcing. The amplitude of the oscillations determines which
periodic orbit the pendulum is in, and there are inﬁnitely many possibilities.
Each orbit is stable, because if we give the pendulum a tiny push, it will settle
into a new periodic orbit with a slightly different amplitude. In other words, any
small perturbation of our original periodic orbit will shift us to a “nearby” orbit.
None of these periodic orbits is asymptotically stable, because any tiny push we
apply to the pendulum will land us on a different periodic orbit than the one we
started with (since the pendulum is frictionless).
Example 4.1.6. Consider the planar system in polar coordinates given by
dr
dt
= r(1 −r) and
dθ
dt
= 1.
To envision the motion along a trajectory in the phase plane, notice that our
angular variable θ continually increases with speed 1, while the radial variable
obeys a logistic differential equation. By itself, the equation
dr
dt
= r(1 −r) has
two equilibria: r = 0 and r = 1. The “Jacobian” for this 1d equation is simply
given by the derivative J f (r) = 1 −2r. For r = 0, we have J f (0) = 1 > 0, which
128 peri odi c orbi ts and the poi ncaré bendi xon theorem
x
2
x
1
Figure 4.3. Phase portrait for the planar system
dr
dt
= r(1 −r),
dθ
dt
= 1.
means that r = 0 is an unstable equilibrium. On the other hand, J f (1) = −1 < 0
means that r = 1 is a stable equilibrium. The graph of r = 1 is a circle of radius
1, and in this case it corresponds to an asymptotically stable periodic orbit. Also,
r = 0 corresponds to the origin, which is an unstable equilibrium. A sketch of
the phase portrait for this system appears in Figure 4.3.
Example 4.1.7. A similar argument shows that the planar system
dr
dt
= r(r −1) and
dθ
dt
= 1.
has an unstable periodic orbit corresponding to the circular trajectory r = 1.
Deﬁnition 4.1.8. A periodic orbit is called a limit cycle if there exists a point x
0
that does not lie on the periodic orbit but such that the trajectory φ
t
(x
0
) converges
to the periodic orbit either as t →∞ or as t → −∞.
In the above two examples, r = 1 corresponds to a limit cycle solution. In
general, asymptotically stable periodic orbits and unstable periodic orbits are
examples of limit cycle solutions. On the other hand, the constantcoefﬁcient
system x
/
= Ax where
A =
_
0 −1
1 0
_
does not have any limit cycle solutions although it does have inﬁnitely many
periodic solutions. The phase portrait for that linear system consists of a family
of concentric circles.
peri odi c, heterocli ni c, and homocli ni c orbi ts 129
Before concluding our discussion of periodic orbits, we remark that the
PoincaréBendixon theorem ensures that planar systems x
/
= f (x) (with f
continuously differentiable) can never exhibit behavior that is too “crazy”. Specif
ically, such planar systems can never have chaotic solutions (for a proof, see
Hubbard and West [5]). We now state the deﬁnition of chaos as it appears in
Strogatz [11].
Deﬁnition 4.1.9. Chaos is aperiodic behavior in a deterministic system which
exhibits sensitive dependence on initial conditions.
There are three important phrases in the deﬁnition of chaos. By aperiodic
behavior, we mean that there are solutions which do not converge to a limit cycle
or stable equilibrium. Deterministic means that the aperiodic behavior is not a
result of “cheating” by artiﬁcially incorporating randomness into the system.
Sensitive dependence on initial conditions means that slight changes in initial
conditions can cause huge differences in how solutions behave.
Example 4.1.10. The Lorenz equations are given by
dx
dt
= σ(y −x)
dy
dt
= rx −y −xz
dz
dt
= xy −bz,
(4.1)
where σ, r, and b are constants. Clearly the origin is an equilibrium solution. It
is possible to show that if 0 < r < 1, then the origin is globally asymptotically
stable. This is accomplished by using V(x, y, z) = rx
2
+σy
2
+σz
2
as a Lyapunov
function. Moreover, if r > 1, you can show that there is a onedimensional
unstable manifold at the origin.
For the speciﬁc choices σ = 10, r = 28 and b = 8/3, this system is known to
exhibit chaos. A graph of a trajectory in the phase space is shown in Figure 4.4.
Although the trajectory remains conﬁned to a bounded set, the solution is
aperiodic and never converges to an equilibrium solution or limit cycle.
As we mentioned above, the PoincaréBendixon Theorem rules out chaos for
planar systems. Apart from equilibria and limit cycle solutions, there are very
few other types of interesting behavior that planar systems can exhibit. Two
exceptions are discussed in the next section.
130 heterocli ni c and homocli ni c orbi ts
z
x
y
Figure 4.4. Chaotic solution of the Lorenz equations (4.1).
4.2. Heteroclinic and Homoclinic Orbits
Besides periodic solutions and equilibria, can there be nonconstant solutions
of x
/
= f (x) which remain bounded? The answer is yes, as we illustrate via the
following examples.
A heteroclinic orbit. Consider the secondorder ode
u
//
+ cu
/
−u(u − β)(u −1) = 0, (4.2)
where c and β are positive constants and 0 < β <
1
2
. Introducing w = u
/
, we
write the equation as a system of two ﬁrstorder equations
u
/
= w
w
/
= u(u − β)(u −1) −cw = u
3
−(1 + β)u
2
+ βu −cw.
(4.3)
Note that there are three equilibria for this system. Setting u
/
= 0 forces w = 0,
and then setting w
/
= 0 forces u = 0, u = β, or u = 1. The Jacobian matrix for
this system is
J f (u, w) =
_
0 1
3u
2
−2(1 + β)u + β −c
_
.
peri odi c, heterocli ni c, and homocli ni c orbi ts 131
At the equilibrium (u, w) = (0, 0), the Jacobian is
J f (0, 0) =
_
0 1
β −c
_
,
which has characteristic equation λ
2
+ cλ − β = 0. Since the determinant
of J f (0, 0) is −β < 0, we immediately conclude that (0, 0) is a hyperbolic
equilibrium and that there is a saddle at the origin. The eigenvalues of J f (0, 0)
are λ
±
=
1
2
_
−c ±
_
c
2
+4β
_
, the larger of which is positive and the smaller of
which is negative. To calculate an eigenvector for the positive eigenvalue λ
+
, we
form the matrix
A −λ
+
I =
_
−λ
+
1
β −c −λ
+
_
which rowreduces to
_
−λ
+
1
0 0
_
.
Eigenvectors v must satisfy −λ
+
v
1
+ v
2
= 0, implying that v
2
= λ
+
v
1
. It is
convenient to treat v
1
as our free variable so that
_
v
1
v
2
_
=
_
v
1
λ
+
v
1
_
= v
1
_
1
λ
+
_
.
It follows that
_
1
λ
+
_
is an eigenvector for λ
+
, and a similar calculation reveals that
_
1
λ
−
_
is an eigenvector for the negative eigenvalue λ
−
. The spans of these two eigen
vectors form the unstable and stable subspaces, and since λ
−
< 0 < λ
+
the
orientation of these vectors is as sketched in Figure 4.5. The same sorts of cal
culations show that the equilibrium (u, w) = (1, 0) is also a saddle and has a
similar orientation as the saddle at the equilibrium (Figure 4.5).
132 heterocli ni c and homocli ni c orbi ts
The behavior at the equilibrium (u, w) = (β, 0) is a bit different. The Jacobian
matrix is
J f (β, 0) =
_
0 1
β
2
− β −c
_
,
and we note that the determinant β − β
2
is positive because 0 < β < 1. Hence,
this equilibrium is not a saddle like the other two. In fact, the eigenvalues
−c ±
_
c
2
−4(β − β
2
)
2
both have negative real part. It follows that this equilibrium is either a stable
node or a stable focus, depending upon the relative sizes of the constants β and
c.
Interesting fact: If the constant c is chosen appropriately, we can force the stable
manifold W
s
(0, 0) at the origin to coincide with the unstable manifold W
u
(1, 0)
of the equilibrium (1, 0). This forms a special trajectory in the phase plane which
connects the two equilibria, and there is a name for this type of solution.
Deﬁnition 4.2.1. Suppose that x
∗
and x
∗∗
are two distinct equilibria of the system
x
/
= f (x). A solution x(t) with the property that
lim
t→−∞
x(t) = x
∗
and lim
t→∞
x(t) = x
∗∗
is called a heteroclinic orbit.
In our example above, it is actually possible to use a special trick to ﬁnd
the heteroclinic orbit analytically. Indeed, consider the simpler ode given by
u
/
= au(u −1), where a > 0 is a constant. Unlike our original system, this ode
has only two equilibria: u = 0 and u = 1. It is easy to check that 0 is stable
and 1 is unstable. We claim that for special choices of the constants a and c, the
solutions of this simpler differential equation are also solutions of the original
secondorder equation (4.2). To see this, we will substitute u
/
= au(u −1) into
Equation (4.2). Since u
/
= a(u
2
−u), we calculate that
u
//
= a(2uu
/
−u
/
) = a(2u −1)u
/
= a
2
u(u −1)(2u −1).
peri odi c, heterocli ni c, and homocli ni c orbi ts 133
E (0,0)
s
E (0,0)
u
E (1,0)
s
E (1,0)
u
u
w
β 1
Figure 4.5. Two saddle equilibria for Equations (4.3).
Equation (4.2) becomes
a
2
u(u −1)(2u −1) + cau(u −1) −u(u − β)(u −1) = 0
which, after factoring out u(u −1), can be rewritten as
u(u −1)
_
(2a
2
−1)u + (β + ca −a
2
)
_
= 0.
The only way the left hand side, a function of u, could be identically equal to 0
is if both
2a
2
−1 = 0 and β + ca −a
2
= 0.
The solution of this system of two equations is
a =
1
√
2
and c =
√
2
_
1
2
− β
_
.
Note that c > 0 since 0 < β <
1
2
. In summary, for these special choices of
c and a, the solution of the equation u
/
= au(u −1) is also a solution of the
original equation (4.2). It is straightforward to solve this equation by separation
of variables. The corresponding trajectory in the uwphase plane is a heteroclinic
orbit. As t → −∞, the trajectory connects to the equilibrium (u, w) = (1, 0), and
as t →∞, the trajectory connects to the equilibrium (u, w) = (0, 0). The solution
exists for all time t, is bounded, and is neither an equilibrium nor a periodic orbit
134 heterocli ni c and homocli ni c orbi ts
u
w
β
1 t
u
0
1
Figure 4.6. A heteroclinic orbit for Equation (4.2). The left panel shows
the uwphase plane, and the heteroclinic orbit appears in bold. The
right panel shows a graph of u versus t. Note that u → 0 as t →∞ and
u → 1 as t → −∞.
(see Figure 4.6). The heteroclinic orbit forms a connection of the stable manifold
W
s
(0, 0) with the unstable manifold W
u
(1, 0).
A homoclinic orbit. We now discuss a special type of orbit in which the stable
and unstable manifolds of the same equilibrium are connected. The example we
give is that of a double well potential. More exactly, consider the system
x
/
= y and y
/
= x −x
3
, (4.4)
which has three equilibria. Linearization indicates that (x, y) = (0, 0) is a
saddle equilibrium. The equilibria (1, 0) and (−1, 0) are nonhyperbolic, and
are centers for the linearized systems. It is actually possible to show that they
are centers for the nonlinear system as well, by deﬁning the energy functional
E(x, y) =
1
2
y
2
−
1
2
x
2
+
1
4
x
4
and arguing that E(x, y) remains constant along phase
plane trajectories. Figure 4.7 shows a rough sketch of the phase plane for this
system. Notice that the unstable manifold W
u
(0, 0) at the origin happens to
coincide with the stable manifold W
s
(0, 0). This closed loop trajectory is not
a periodic orbit—a sketch of the corresponding solution x(t) also appears in
Figure 4.7. This special orbit is called a homoclinic orbit.
peri odi c, heterocli ni c, and homocli ni c orbi ts 135
x
t
y
x
Figure 4.7. Two homoclinic orbits for Equation (4.4). The left panel
shows the xyphase plane, and the homoclinic orbits appear in bold. The
right panel shows a graph of x versus t for one of the two homoclinic
orbits. Note that x → 0 as t → ±∞.
Exercises
1. This series of questions concerns the famous Lorenz system
dx
dt
= −σx + σy
dy
dt
= rx −y −xz
dz
dt
= xy − βz,
where σ, r, and β are positive parameters. The Lorenz system exhibits chaos if
σ = 10, r = 28, and β = 8/3. Below, you will learn why chaos could never
occur if r < 1.
(a) Explain why, independent of the choice of initial conditions, the initial
value problem for the Lorenz system is guaranteed to have a unique
solution (at least locally).
(b) Show that the origin is a locally asymptotically stable equilibrium if 0 <
r < 1.
(c) Show that if r > 1, then there will be a onedimensional unstable manifold
at the origin.
(d) The origin is actually globally asymptotically stable if r < 1, and the
remaining parts of this exercise will guide you through a proof. To start,
136 heterocli ni c and homocli ni c orbi ts
deﬁne the function V(x, y, z) =
1
σ
x
2
+ y
2
+ z
2
. Show that
∇V • f = −2
_
x −
_
r +1
2
_
y
_
2
−2
_
1 −
_
r +1
2
_
2
_
y
2
−2βz
2
,
where f denotes the right hand side of the Lorenz system.
(e) Show that if r < 1, then
1 −
_
r +1
2
_
2
> 0.
Then, explain why this implies that ∇V • f is strictly less than 0 (except at
the origin) whenever r < 1.
(f) Finally, explain why V is a Lyapunov function and why you can conclude
that the origin is a global attractor if r < 1.
2. Consider the system
dx
dt
= −y −x(1 −x
2
−y
2
)
dy
dt
= x −y(1 −x
2
−y
2
).
Show that this system has an unstable periodic solution and carefully sketch
the phase portrait.
3. Consider the system
dx
dt
= −y + x(4 −x
2
−y
2
)
dy
dt
= x + y(9 −x
2
−y
2
).
You may assume that the origin is the only equilibrium of this system. Classify
the local stability of the origin. Then, show that this system has at least one
stable, periodic solution.
4. Consider the system
dx
dt
= −y + x(r
2
−6r +8)
dy
dt
= x + y(r
2
−6r +8),
peri odi c, heterocli ni c, and homocli ni c orbi ts 137
where r
2
= x
2
+ y
2
. Use the PoincaréBendixon Theorem to prove that this
system has both stable and unstable periodic orbits by following these steps:
(a) Show that the origin is the only equilibrium of this system.
(b) Using the chain rule, differentiate both sides of r
2
= x
2
+ y
2
with respect
to t. Then, assuming r ,= 0 (i.e., excluding the equilibrium solution), solve
for
dr
dt
. You should obtain an autonomous ode for r.
(c) Using the equation for
dr
dt
you found in Part (b), show that
dr
dt
> 0 on the
circle r = 1 and that
dr
dt
< 0 on the circle r = 3. Use the PoincaréBendixon
Theorem to conclude that there is at least one stable periodic orbit within
the annulus 1 < r < 3.
(d) Using the equation for
dr
dt
you found in Part (b), show that
dr
dt
> 0 on the
circle r = 5. Combined with the fact that
dr
dt
< 0 on the circle r = 3, this
seems to suggest that an unstable periodic orbit exists inside the annulus
3 < r < 5. To prove this, make the substitution t → −t, which “reverses”
the ﬂow of time. Then, use the PoincaréBendixon Theorem to show
that the resulting system has a stable periodic orbit inside the annulus
3 < r < 5. Finally, conclude that the original system (i.e., going forward
in time) has an unstable periodic orbit inside that annulus.
5. Here is an instance of the famous FitzHughNagumo nerve membrane model:
dx
dt
= −x(x −1)(x +1) −y
dy
dt
= x −
1
2
y.
(a) Show that this system has exactly one equilibrium and that it is unstable.
(b) Consider the rectangular path Γ with corners (
√
3, 2
√
3), (−
√
3, 2
√
3),
(−
√
3, −2
√
3), and (
√
3, −2
√
3) as illustrated in Figure 4.8. By showing
that the ﬂow is directed inward on this path, use the PoincaréBendixon
Theorem to conclude that these equations have at least one periodic
solution. Hint: You will need to parametrize each edge of the rectangle
separately, and there are many possible parameterizations.
6. Consider the secondorder nonlinear ode
u
//
+
5
√
6
u
/
+ u(1 −u) = 0.
138 heterocli ni c and homocli ni c orbi ts
y
x
Figure 4.8. Rectangular path to use in exercise involving FitzHugh
Nagumo model.
(a) Write the ode as a system of two ﬁrstorder odes by setting v = u
/
. Show
that the resulting system has two equilibria, and that one equilibrium is a
stable node while the other is a saddle.
(b) Verify that
u(t) =
1
_
1 + (
√
2 −1)e
t/
√
6
_
2
is a solution of the ode and that it corresponds to a heteroclinic orbit.
Plot u(t) versus t and sketch the corresponding trajectory in the uvphase
plane.
7. In this exercise, you will ﬁnd a homoclinic orbit for a particular ode. Consider
the secondorder, nonlinear equation
v
//
+3v
2
−σv = 0
where σ is a constant. This ode could be written as a system in the usual
way, by letting w = v
/
. We claim that there is a nonzero solution v(t) which
has the property that v, v
/
, v
//
→ 0 as t → ±∞. This solution, if plotted in the
vwphase plane, would correspond to a homoclinic orbit.
(a) Multiply both sides of the above ode by v
/
and show that the resulting
equation is equivalent to
d
dt
(v
/
)
2
2
=
d
dt
_
−v
3
+
σ
2
v
2
_
.
peri odi c, heterocli ni c, and homocli ni c orbi ts 139
(b) Now integrate both sides of the equation in Part (a), and let β denote the
integration constant. Explain why we may assume that β = 0.
(c) By algebraically solving for v
/
, show that
dv
dt
= ±v
√
σ −2v.
Then, taking the minus sign for later convenience, use separation of
variables to show that
t = −
_
dv
v
√
σ −2v
.
(d) Recall the following deﬁnitions of hyperbolic trigonometric functions:
sinh(z) =
1
2
(e
z
−e
−z
) cosh(z) =
1
2
(e
z
+e
−z
)
tanh(z) =
sinh(z)
cosh(z)
sech(z) =
1
cosh(z)
In the integral you wrote down in Part (c), make the (hyperbolic) trigono
metric substitution
v =
σ
2
sech
2
(θ) dv = −σsech
2
(θ) tanh(θ) dθ
and note that
v
√
σ −2v =
σ
3/2
2
sech
2
(θ) tanh(θ).
Do the resulting integral to show that
t =
2
√
σ
θ + C
where C is an integration constant.
(e) Use algebra to show that
v(t) =
σ
2
sech
2
_√
σ
2
(t −C)
_
.
Then, plot the function v(t) versus t for the special case σ = 1 and C = 0
to get a sense of what the homoclinic solution would look like. Finally, set
w = v
/
and sketch the homoclinic trajectory in the vwphase plane.
CHAPTER 5
Bifurcations
I
n practice, we often deal with odes which contain parameters (unspeciﬁed
constants) whose values can profoundly inﬂuence the dynamical behavior of
the system. For example, suppose we model population changes for a species.
The birth and death rates of the species would be examples of parameters whose
values would substantially impact that behavior of solutions of the underlying
differential equation.
Example 5.0.2. Consider the ode
dx
dt
= µx, where µ is a parameter. The solution
of this equation is x(t) = x
0
e
µt
, where x
0
= x(0). Notice that if µ > 0, the
solutions exhibit exponential growth, whereas if µ < 0 we observe exponential
decay. If µ = 0, solutions are constant. The critical value µ = 0 marks the
“boundary” between two very different types of dynamical behavior.
Deﬁnition 5.0.3. Consider a system of odes of the form x
/
(t) = f (x; µ),
where µ is a parameter. A bifurcation is any major qualitative change in the
dynamical behavior of the system in response to varying the parameter µ.
In the previous example, we would say that a bifurcation occurs at µ = 0,
because the equilibrium x = 0 changes from stable (µ < 0) to unstable
(µ > 0).
5.1. Three Basic Bifurcations
There are many ways that the qualitative behavior of a system can be drastically
altered in response to changes in parameters. Equilibria and/or periodic solu
tions can be created or destroyed, or they can change their stability. In what
follows, we will survey several common types of bifurcations. For simplicity, we
140
bi furcati ons 141
will deal primarily with odes with only one dependent variable. An excellent
reference for this material is provided by Strogatz [11].
Saddlenode bifurcations. First, we discuss one common mechanism for the
birth or destruction of an equilibrium solution. The canonical example to keep
in mind is given by the ode
dx
dt
= µ + x
2
. (5.1)
Suppose µ < 0. The equilibria satisfy x
2
+µ = 0, which implies that x = ±
√
−µ
are the two equilibria of the system. To determine whether they are stable,
we compute that the “Jacobian” of f (x) = x
2
+ µ is simply f
/
(x) = 2x. By
calculating f
/
(
√
−µ) = 2
√
−µ > 0, we ﬁnd that
√
−µ is an unstable equilibrium.
Similarly, since f
/
(−
√
−µ) = −2
√
−µ < 0, it follows that −
√
−µ is a stable
equilibrium. Next, suppose that µ = 0. There is only one equilibrium, namely
x = 0. Although it is nonhyperbolic, it is easy to check (via separation of
variables) that this equilibrium is unstable. Finally, suppose µ > 0. Then
x
2
+ µ > 0 for all x, implying that there are no equilibria. In summary, as µ
increases from negative to positive, two equilibria (one stable and one unstable)
merge and annihilate each other, leaving no equilibria if µ > 0. Clearly µ = 0 is a
bifurcation point, and this type of bifurcation is called a saddlenode bifurcation. A
minor adjustment of Equation (5.1) reveals that saddlenode bifurcations can also
create new equilibria as the parameter µ increases. The equation x
/
= −µ + x
2
experiences a saddlenode bifurcation at µ = 0, creating two equilibria for µ > 0.
Bifurcation diagrams. One very common and useful way to visualize bifurca
tions of a system x
/
= f (x; µ) is to sketch a graph of the equilibrium values of x
as a function of the bifurcation parameter µ. As an illustration, the bifurcation di
agram for Equation (5.1) appears in Figure 5.1. By convention, unstable equilibria
are plotted as dashed curves and stable equilibria are plotted as solid curves. In
Figure 5.1, the dashed curve corresponds to the unstable equilibrium x =
√
−µ
and the solid curve corresponds to the stable equilibrium x = −
√
−µ. The
bifurcation diagram allows us to visualize the qualitative behavior of solutions
for various choices of initial conditions and the parameter µ. For example, if
we start at the initial condition x
0
in Figure 5.1, the unstable equilibrium will
repel us and x will decrease towards the stable equilibrium (as indicated by an
142 three basi c bi furcati ons
x
o
unstable
stable
x
µ
no
equilibria
Figure 5.1. Saddlenode bifurcation in Equation (5.1). For µ < 0, there
are two equilibria: x = ±
√
−µ. The larger of these is unstable (dashed
curve) and the smaller is stable (solid curve). For µ > 0, there are no
equilibria.
arrow in the Figure). The bifurcation diagram for x
/
= −µ + x
2
can be obtained
by reﬂecting the parabola in Figure 5.1 across the vertical axis.
Transcritical bifurcations. The saddlenode bifurcation is a very common way
for equilibria to be created or destroyed as a parameter µ is varied. We now
describe a mechanism by which two equilibria can exchange their stability. The
canonical example of a transcritical bifurcation is given by the equation
dx
dt
= µx −x
2
. (5.2)
Setting
dx
dt
= 0, we ﬁnd that there are two equilibria: x = 0 and x = µ. Note
that the former is an equilibrium independent of the choice of µ. To check the
stability of these equilibria, we compute that the “Jacobian” of f (x) = µx −x
2
is simply f
/
(x) = µ −2x. Since f
/
(0) = µ, we conclude that x = 0 is stable if
µ < 0 and unstable if µ > 0. Similarly, since f
/
(µ) = −µ, we conclude that x = µ
is stable if µ > 0 and unstable if µ < 0. In this case, µ = 0 is the bifurcation
point, and the two equilibria exchange their stability there. The corresponding
bifurcation diagram appears as Figure 5.2, and the arrows in the ﬁgure indicate
how x would change starting from various choices of x(0) and µ. For example, if
µ > 0 and we choose any initial condition x
0
∈ (0, µ), then x(t) → µ as t →∞.
bi furcati ons 143
x
µ
Figure 5.2. Transcritical bifurcation in Equation (5.2). For µ < 0, there
are two equilibria: x = 0 is stable (solid) and x = µ is unstable (dashed).
A bifurcation occurs at µ = 0, and the two equilibria exchange stability.
Arrows indicate how x(t) will change as t increases, starting from
various choices of µ and initial conditions x(0).
Example 5.1.1. Equation (5.2) is similar to a wellknown logistic model for
population growth
x
/
= µx(M−x), (5.3)
where x represents population of a species, M is the maximum population that
the environment can sustain, and the parameter µ measures the birth rate minus
the death rate. If µ ,= 0, the equation has exactly two equilibria: x = 0 (extinction)
and x = M (proliferation), neither of which depend on µ. If µ < 0, the death
rate exceeds the birth rate and x = 0 is a stable equilibrium while x = M is
unstable. If µ > 0, then x = M is stable and x = 0 is unstable, and we expect
the population to approach the environment’s carrying capacity as t →∞. The
transcritical bifurcation at µ = 0 marks the “threshold” between the species’
extinction and proliferation. Figure 5.3 shows the bifurcation diagram for (5.3)
(compare to Figure 5.2).
Pitchfork bifurcations. One type of bifurcation that is common in physical
problems involving some sort of symmetry is the pitchfork bifurcation. As an
example, suppose that we balance a weight on top of a vertical beam (Figure 5.4).
If the weight is small, the system is in a stable equilibrium. If the weight increases
beyond a certain point, the equilibrium loses its stability. Any slight deﬂection of
the beam from a perfectly vertical position will cause the beam to buckle. If we
144 three basi c bi furcati ons
x
M
0 extinction
proliferation
µ
Figure 5.3. Transcritical bifurcation in Equation (5.3). For µ ,= 0, there
are two equilibria: x = 0 and x = M. The bifurcation at µ = 0 causes
the two equilibria to exchange stability. Arrows indicate how x(t) will
change as t increases, starting from various choices of µ and initial
conditions x(0).
regard this system as “twodimensional”, then the beam will either buckle left or
right. Another example is provided by a bead on a rotating hoop (Figure 5.5).
For slow rotation speeds, the bead rests at the bottom of the hoop. However, this
equilibrium loses its stability if the rotation speed is increased beyond a certain
critical value. In these ﬁrst of these examples, the bifurcation parameter µ would
be the weight of the object being supported by the beam. In the latter example,
the bifurcation parameter µ is the speed of rotation of the hoop.
The canonical example of a pitchfork bifurcation is provided by the equation
dx
dt
= µx −x
3
. (5.4)
The right hand side of this equation can be factored as f (x; µ) = x(µ − x
2
).
If µ < 0, then the quadratic factor has no real roots and therefore x = 0 is the
only equilibrium. On the other hand, if µ > 0, then further factorization reveals
that f (x; µ) = x(
√
µ − x)(
√
µ + x). Thus, for positive µ there are 3 equilibria:
x = 0 and x = ±
√
µ. Using the same sort of stability analysis we used in our
discussions of saddlenode and transcritical bifurcations, you should convince
yourself that x = 0 is stable if µ < 0 and unstable if µ > 0. Both x = ±
√
µ
are stable equilibria for µ > 0. The bifurcation diagram in Figure 5.6 should
convince you that the name pitchfork bifurcation is indeed appropriate. As we
bi furcati ons 145
weight
b
e
a
m
Figure 5.4. A weight being supported by a vertical beam.
Figure 5.5. A bead on a rotating hoop. The vertical lines indicate the
axis of rotation. For slow rotation speeds (left panel), the bead rests in
a stable equilibrium at the bottom of the hoop. If the rotation speed
is increased, this equilibrium loses stability and the bead settles into a
new stable position (right panel).
see in the Figure, as µ increases from negative to positive, an equilibrium loses
stability at the same instant that two new equilibria are born.
Subcritical versus supercritical bifurcations. One further way of classifying
bifurcations is according to whether an equilibrium gains or loses stability
as the parameter µ is increased. The pitchfork bifurcation in Equation (5.4)
above is considered to be a supercritical bifurcation. An example of a subcritical
146 three basi c bi furcati ons
x
µ
Figure 5.6. Pitchfork bifurcation in Equation (5.4). For µ < 0, there is
only one equilibrium (x = 0) and it is stable. For µ > 0, there are three
equilibria: x = 0 is unstable and x = ±
√
µ are stable. Arrows indicate
qualitative behavior of solutions starting from various initial conditions
and choices of the parameter µ.
pitchfork bifurcation would be given by the equation x
/
= µx + x
3
. The reader
is encouraged to sketch a bifurcation diagram for this equation. You should
ﬁnd that x = 0 is a stable equilibrium for µ < 0 and is unstable if µ > 0. For
µ < 0, there are two other equilibria x = ±
√
−µ, both of which are unstable. In
what follows, we will rarely distinguish between subcritical and supercritical
bifurcations.
Example 5.1.2. Let µ be a real parameter. Sketch the bifurcation diagram for the
ode
dx
dt
= f
µ
(x) = 5 −
µ
1 + x
2
. (5.5)
Solution: Equilibria would satisfy the equation
5 −
µ
1 + x
2
= 0,
which can be manipulated to reveal that x
2
= µ/5 −1. There are no equilibria
for µ < 5, but if µ > 5 then there are two equilibria:
x = ±
_
µ
5
−1.
bi furcati ons 147
equilibria
no
x
unstable
stable
µ
5
Figure 5.7. Bifurcation diagram for Equation (5.5). A saddlenode
bifurcation occurs at µ = 5.
We can already conclude that this is an example of a saddlenode bifurcation. It
remains to determine which equilibria are stable. Taking the derivative of f
µ
(x),
we have
f
/
µ
(x) =
2µx
(1 + x
2
)
2
.
It follows that
f
/
µ
(
_
µ/5 −1) =
2µ
_
µ/5 −1
(µ/5)
2
,
which is positive. It follows that for µ > 5, the larger of the two equilibria is
unstable. The other ﬁxed point is stable. The bifurcation diagram appears in
Figure 5.7.
Example 5.1.3 (Courtesy of S. Lazaryan). Let µ be a real parameter. Sketch a
bifurcation diagram for the ode
dx
dt
= f
µ
(x) = x(x
2
−µ)
_
e
x
2
−µ
_
. (5.6)
Solution: Equilibria must satisfy the equation
x(x
2
−µ)
_
e
x
2
−µ
_
= 0,
and certainly x = 0 is an equilibrium regardless of the value of µ. If µ > 0, then
the factor x
2
−µ has two real roots, x = ±
√
µ. In order for e
x
2
−µ = 0 to have
real solutions, we would need µ ≥ 1. If µ > 1, algebra reveals that x = ±
_
ln µ
are also equilibria. In total, there are ﬁve equilibria if µ > 1.
148 dependence of soluti ons on parameters
To test the stability of these equilibria, we must compute the “Jacobian” of
f
µ
(x), taking the derivative with respect to x. Expanding f
µ
(x) as
f
µ
(x) = x
3
e
x
2
−µxe
x
2
−µx
3
+µ
2
x,
we use the chain and product rules to calculate
f
/
µ
(x) = 3x
2
e
x
2
+2x
4
e
x
2
−µe
x
2
−2µx
2
e
x
2
−3µx
2
+µ
2
.
For the equilibrium x = 0, we have f
/
µ
(0) = −µ +µ
2
= µ(µ −1). This quantity
is negative if 0 < µ < 1 and positive otherwise. We conclude that x = 0 is a
stable equilibrium if 0 < µ < 1 and is unstable otherwise.
Next, assume µ > 0 and consider the pair of equilibria x = ±
√
µ. Notice that
x
2
= µ and x
4
= µ
2
for both of these equilibria. Using the above formula for
f
/
µ
(x) we have
f
/
µ
(±
√
µ) = 3µe
µ
+2µ
2
e
µ
−µe
µ
−2µ
2
e
µ
−3µ
2
+µ
2
= 2µe
µ
−2µ
2
= 2µ (e
µ
−µ) .
Since e
µ
> µ, we conclude that f
/
µ
(±
√
µ) > 0 for all positive µ. This implies that
the two equilibria x = ±
√
µ are unstable for µ > 0.
Finally, assume µ > 1 and consider the pair of equilibria x = ±
_
ln µ. In this
case, we have x
2
= ln µ and e
x
2
= µ. Our formula for f
/
µ
(x) yields
f
/
µ
(±
_
ln µ) = 3µ ln µ +2µ(ln µ)
2
−µ
2
−2µ
2
ln µ −3µ ln µ +µ
2
= 2µ ln µ (ln µ −µ) .
Examining the factors individually, note that (ln µ −µ) < 0 and ln µ > 0 for
µ > 1. It follows that f
/
µ
(±
_
ln µ) < 0 for all µ > 1, implying that these two
equilibria are stable. A bifurcation diagram summarizing the stability of the
equilibria appears in Figure 5.8.
5.2. Dependence of Solutions on Parameters
In a previous chapter, we discussed how changes in initial conditions can affect
how solutions behave. We even derived an estimate for how fast solution
bi furcati ons 149
1 0
x
µ
Figure 5.8. Bifurcation diagram for Equation (5.6). A pitchfork bifur
cation occurs at µ = 0, creating two unstable equilibria x = ±
√
µ and
changing x = 0 into a stable equilibrium. Another pitchfork bifurca
tion occurs at µ = 1, destabilizing x = 0 and creating two new stable
equilibria, x = ±
_
ln µ.
trajectories can separate if we start from two different “nearby” initial conditions.
We now perform the same sort of calculation to determine how changes in a
parameter µ can affect the behavior of solutions.
Consider the two differential equations
x
/
= f (x; µ) and y
/
= f (y; λ), (5.7)
and assume that both have the same initial condition: x(0) = a = y(0). In other
words, we are considering two copies of the same initial value problem, but
with two different parameter choices, µ and λ. Assume that the function f is
continuously differentiable with respect to both arguments. From a lemma that
we proved when analyzing dependence of solutions on initial conditions, there
exist positive constants K
1
and K
2
such that
 f (x; µ) − f (x; λ) ≤ K
1
µ −λ and  f (x; λ) − f (y; λ) ≤ K
2
x −y.
Let K = max{K
1
, K
2
}. Then we can use the triangle inequality to estimate the
difference between the right hand sides of the odes in Equation (5.7):
 f (x; µ) − f (y; λ) =  f (x; µ) − f (x; λ) + f (x; λ) − f (y; λ)
≤  f (x; µ) − f (x; λ) +  f (x; λ) − f (y; λ)
150 dependence of soluti ons on parameters
≤ K
1
µ −λ + K
2
x −y ≤ K (µ −λ + x −y) .
We are now in a position to measure the gap between the solutions of the two
different initial value problems. Writing them as integral equations,
x(t) = a +
_
t
0
f (x(s); µ) ds and y(t) = a +
_
t
0
f (y(s); λ) ds.
Using our above inequality, we now make the estimate
y(t) −x(t) =
¸
¸
¸
¸
_
t
0
f (y(s); λ) ds −
_
t
0
f (x(s); µ) ds
¸
¸
¸
¸
≤
_
t
0
 f (y(s); λ) − f (x(s); µ) ds ≤
_
t
0
K (µ −λ + x −y) ds
= K
_
t
0
µ −λ ds + K
_
t
0
y(s) −x(s) ds.
In the last line, the ﬁrst integrand is simply a constant. If we introduce the
abbreviation B = Kµ −λ and actually evaluate the integral, our overall estimate
has been written as
y(t) −x(t) ≤ Bt + K
_
t
0
y(s) −x(s) ds.
This is set up perfectly for the Strong Gronwall Inequality 3.3.12, and we conclude
that
y(t) −x(t) ≤
B
K
_
e
Kt
−1
_
= µ −λ
_
e
Kt
−1
_
.
Consequence: Suppose the function f is wellbehaved in the sense that there
exists a global constant K such that
 f (x; µ) − f (y; λ) ≤ K (µ −λ + x −y) (5.8)
for all real x, y, µ, and λ. Then the solutions of the two initial value problems
(corresponding to different choices of the parameter) can separate at most ex
ponentially fast. This ensures that, at least for wellbehaved problems of this
sort, any bifurcations caused by changing the parameter values cannot cause a
“catastrophic” change in solution behavior. For example, if solutions do not blow
up to ∞ in ﬁnite time when we use the parameter µ, then they cannot blow up
in ﬁnite time if we switch our parameter value from µ to λ.
bi furcati ons 151
Warning: If we have a system x
/
= f (x; µ) for which f does not obey inequal
ity (5.8) globally, then catastrophes can happen. For example, recall our canonical
example of a saddlenode bifurcation:
dx
dt
= µ + x
2
. Suppose µ < 0 and that
the initial condition x
0
satisﬁes the inequality −
√
−µ < x
0
<
√
−µ. Then the
solution x(t) will converge to the stable equilibrium −
√
−µ as t → ∞. Now
suppose we use the same initial condition, but change the parameter µ to some
positive number. Then the solution of the initial value problem will blow up
to ∞ in ﬁnite time! The problem is that for the function f (x; µ) = µ + x
2
, there
is no global constant K for which inequality (5.8) will hold. More exactly, it is
impossible to ﬁnd one special, positive number K for which the inequality
x
2
−y
2
 ≤ Kx −y
for all x, y in R.
5.3. AndronovHopf Bifurcations
Up to now, we have considered bifurcations in systems with a single dependent
variable. Of course, this is very restrictive because there is precisely one real
“eigenvalue” associated with such systems. Higherdimensional systems can
experience other types of bifurcations that our onedimensional systems could
not. For example, what are the possible ways that a stable equilibrium of a planar
system could lose its stability as we vary a parameter µ? If both eigenvalues
are real and one of them changes from negative to positive, the equilibrium
would change from a sink (stable node) to a saddle. We could also have two
real eigenvalues simultaneously change from negative to positive, converting the
sink to a source (unstable node). However, there is a new possibility—we could
have a pair of complex conjugate eigenvalues change from negative real part to
positive real part, causing a transition from a stable focus to an unstable focus.
We now provide an example of this sort of bifurcation.
Example 5.3.1. Consider the planar system
dx
dt
= y and
dy
dt
= (µ −x
2
)y −x, (5.9)
where µ is a parameter. It is easily checked that, independent of µ, the origin
(x, y) = (0, 0) is the only equilibrium. The Jacobian matrix associated with this
152 andronov hopf bi furcati ons
0 < µ < 2
y
x x
y
−2 < µ < 0
Figure 5.9. Example of an AndronovHopf bifurcation in Equation (5.9).
The origin is a stable focus (left panel) if −2 < µ < 0 and is an unstable
focus (right panel) if 0 < µ < 2.
system is
J f (x, y) =
_
0 1
−1 −2xy µ −x
2
_
,
from which it follows that
J f (0, 0) =
_
0 1
−1 µ
_
.
The characteristic equation is λ
2
−µλ +1 = 0, and the eigenvalues are
λ
±
=
µ ±
_
µ
2
−4
2
.
Assuming that −2 < µ < 2, these eigenvalues are a complex conjugate pair
with real part µ/2. Notice that the real part is negative if µ < 0 and positive of
µ > 0, which means that the equilibrium (0, 0) changes from a stable focus to
an unstable focus if µ increases from negative to positive. Figure 5.9 shows how
trajectories in the phase plane reverse their orientation as we pass through the
bifurcation value µ = 0. This sort of bifurcation has a name.
Deﬁnition 5.3.2. When an equilibrium changes its stability because a pair of
complex conjugate eigenvalues experiences a change in the sign of their real part,
we say that an AndronovHopf bifurcation has occurred.
bi furcati ons 153
y
x
Figure 5.10. For µ = 0.3, Equations (5.9) have a limit cycle solution
(bold curve).
In the above example, the AndronovHopf bifurcation at µ = 0 causes a
dramatic change in the dynamical behavior of the system. If −2 < µ < 0,
the origin is a stable focus and, because the equilibrium is hyperbolic, the
linearization gives a good indicator of how solutions of the nonlinear system
behave. It is possible (although difﬁcult) to prove that the origin is actually a
global attractor by drawing appropriately chosen closed contours Γ, ﬁnding the
outward normal vectors n to Γ, and proving the f • n < 0 everywhere on Γ. At
the bifurcation point µ = 0, the origin is a weakly stable equilibrium. Indeed,
the distance from the origin cannot increase as we move along trajectories in the
direction of increasing t:
d
dt
[x
2
(t) + y
2
(t)] = 2xx
/
+2yy
/
= −2x
2
y
2
≤ 0.
Finally, we established that for 0 < µ < 2, the origin is an unstable equilibrium.
However, more can be said. By creative use of the PoincaréBendixon Theorem,
one may construct carefully chosen contours to ﬁnd that a stable limit cycle
exists for 0 < µ < 2. The birth of periodic solutions (limit cycles) as we pass the
bifurcation value µ = 0 is no accident. In fact, when AndronovHopf bifurcations
occur, we shall ﬁnd that periodic solutions are born when an equilibrium loses
its stability in response to changes in µ.
154 andronov hopf bi furcati ons
Example 5.3.3. Consider the system
x
/
= −y + x(µ −x
2
−y
2
)
y
/
= x + y(µ −x
2
−y
2
),
(5.10)
where µ is a real parameter. We claim that, regardless of µ, the origin is the only
equilibrium for this system. To see why, set both x
/
= 0 and y
/
= 0 to get the
two algebraic equations
y = x(µ −x
2
−y
2
) and x = −y(µ −x
2
−y
2
).
Substituting the second equation into the ﬁrst, we ﬁnd that
y = −y(µ −x
2
−y
2
)
2
.
If the nonnegative quantity (µ −x
2
−y
2
)
2
is strictly positive, then the only way
the equation can be satisﬁed is if y = 0. This, in turn, implies that x = 0 as well.
On the other hand, if µ −x
2
−y
2
= 0, then the system (5.10) reduces to x
/
= −y
and y
/
= x. Again, the only possible equilibrium is x = y = 0.
The linearization of (5.10) about the origin is
_
x
/
y
/
_
=
_
µ −1
1 µ
_ _
x
y
_
,
and the eigenvalues of the coefﬁcient matrix are µ ± i. As µ changes sign
from negative to positive, the origin loses its stability via an AndronovHopf
bifurcation.
How does this bifurcation affect the dynamical behavior of the system? To ad
dress this question, we remark that Equations (5.10) take a very convenient form
if we convert to polar coordinates. Letting R
2
= x
2
+ y
2
and θ = arctan(y/x),
we derive differential equations for R and θ as follows. First,
d
dt
R
2
=
d
dt
(x
2
+ y
2
)
bi furcati ons 155
implies that 2RR
/
= 2xx
/
+ 2yy
/
, where primes indicate differentiation with
respect to t. Substituting the right hand sides of equations (5.10) for x
/
and y
/
,
2RR
/
= 2x[−y + x(µ −x
2
−y
2
)] +2y[x + y(µ −x
2
−y
2
)]
= 2x
2
(µ −x
2
−y
2
) +2y
2
(µ −x
2
−y
2
)
= 2(x
2
+ y
2
)(µ −x
2
−y
2
) = 2R
2
(µ −R
2
).
Avoiding the equilibrium solution (R = 0), we have obtained R
/
= R(µ − R
2
),
a differential equation that describes the distance R from the origin. For the
angular variable θ, we calculate
dθ
dt
=
d
dt
arctan
_
y
x
_
=
xy
/
−yx
/
x
2
·
1
1 + (y/x)
2
=
xy
/
−yx
/
x
2
+ y
2
=
x[x + y(µ −x
2
−y
2
)] −y[−y + x(µ −x
2
−y
2
)]
x
2
+ y
2
=
x
2
+ y
2
x
2
+ y
2
= 1.
In summary, converting our original system (5.10) to polar coordinates gives
dR
dt
= R(µ −R
2
) and
dθ
dt
= 1.
These two equations are uncoupled and can be handled separately. Since θ
/
= 1,
the angular variable increases with constant speed, and we move counterclock
wise as we follow trajectories in the xyphase plane. The radial variable R
measures distance from the origin. You can verify that if µ < 0, then R = 0 is a
stable equilibrium (corresponding to the origin) for the equation R
/
= R(µ −R
2
).
There are no other equilibria for the case µ < 0. On the other hand, if µ > 0,
then there are two equilibria: R = 0 and R =
√
µ. (Note that we exclude the
equilibrium R = −
√
µ because R is a nonnegative quantity.) It is easily checked
that R = 0 is unstable for µ > 0, whereas R =
√
µ is stable. In the phase
plane, R =
√
µ is a circle of radius
√
µ, and it corresponds to a stable limit cycle:
x =
√
µ cos t and y =
√
µ sin t. Sketches of the phase plane for µ < 0 and µ > 0
appear in Figure 5.11.
The important thing to notice here is that the AndronovHopf bifurcation
at µ = 0 created a family of periodic (limit cycle) solutions. The amplitude
of these limit cycles is set by
√
µ. If we sketch a bifurcation diagram for the
156 andronov hopf bi furcati ons
y
x
y
x
µ > 0
µ < 0
Figure 5.11. Phase portraits for the system (5.10) for µ < 0 (left panel)
and µ > 0 (right panel). For µ < 0, the origin is a stable focus. For
µ > 0, there is a limit cycle solution (indicated in bold) corresponding
to a circle radius
√
µ.
R
µ
Figure 5.12. Bifurcation diagram for R
/
= R(µ − R
2
). This is a “half
pitchfork” because, since the radial variable R represents distance from
the origin, we must exclude the possibility of negative R.
onedimensional problem R
/
= R(µ − R
2
), we obtain a “halfpitchfork”—the
lower prong of the pitchfork is missing because we excluded negative values
of R. Figure 5.12 shows this bifurcation diagram.
Plotting a bifurcation diagram for the AndronovHopf bifurcation in the above
example is more complicated than creating a bifurcation diagram for a one
parameter system with a single dependent variable. This idea is to give a
graphical rendering of the longterm behavior of solutions as a function of the
parameter µ. To do so, we must create a threedimensional plot illustrating how
the value of µ affects the behavior of x and y as t →∞. The bifurcation diagram
bi furcati ons 157
y
x
stable
equilibrium
stable
unstable
equilibrium
µ
limit cycle
Figure 5.13. Bifurcation diagram for the system (5.10). For µ < 0, the
origin is a stable equilibrium. For each µ > 0, the origin is an unstable
equilibrium but there is a stable limit cycle with amplitude
√
µ.
appears in Figure 5.13. Notice that the family of limit cycle solutions forms a
surface (a paraboloid, to be speciﬁc).
Like other types of bifurcations, AndronovHopf bifurcations can be either
supercritical or subcritical. The above is an example of the supercritical case,
in which stable limit cycles are spawned as µ increases beyond a critical value.
It is also possible to have subcritical AndronovHopf bifurcations in which a
stable equilibrium loses stability as µ increases, and a family of unstable limit
cycle solutions is destroyed in the process. Qualitatively, all AndronovHopf
bifurcations exhibit the sort of behavior exhibited by Equations (5.10). Let β ,= 0
and consider the following linear, constantcoefﬁcient system with a pair of
complex conjugate eigenvalues µ ± βi:
_
x
/
y
/
_
=
_
µ −β
β µ
_ _
x
y
_
=
_
µx − βy
βx +µy
_
.
Now let us introduce nonlinearity be adding functions with “quadratic and
higherorder” terms to the right hand sides of these two equations; i.e.,
x
/
= µx − βy + p(x, y)
y
/
= βx +µy + q(x, y).
(5.11)
158 andronov hopf bi furcati ons
Here, we assume that the functions p(x, y) and q(x, y) are analytic functions
of two variables. That is, p(x, y) and q(x, y) can be represented in terms of
convergent power series expansions
p(x, y) =
∑
i+j≥2
a
ij
x
i
y
j
= (a
20
x
2
+ a
11
xy + a
02
y
2
)
+ (a
30
x
3
+ a
21
x
2
y + a
12
xy
2
+ a
03
y
3
) + · · ·
q(x, y) =
∑
i+j≥2
b
ij
x
i
y
j
= (b
20
x
2
+ b
11
xy + b
02
y
2
)
+ (b
30
x
3
+ b
21
x
2
y + b
12
xy
2
+ b
03
y
3
) + · · ·
Notice that we have only included quadratic and higherorder terms, and the
subscripts used on the coefﬁcients correspond to the exponents on x and y.
Roughly speaking, p and q are nothing more than generalizations of polynomials
in two variables. By design, the system (5.11) has exactly one equilibrium (the
origin), and the linearization about the origin is given by
_
x
/
y
/
_
=
_
µ −β
β µ
_ _
x
y
_
.
If µ = 0, then the origin is a center for the linearized system.
In preparation for stating a criterion for the occurrence of an AndronovHopf
bifurcation in (5.11), we will need the following deﬁnition.
Deﬁnition 5.3.4. For the system (5.11), the Lyapunov number is deﬁned as
σ =
3π
2β
_
3(a
30
+ b
03
) + (a
12
+ b
21
) −
2
β
(a
20
b
20
−a
02
b
02
)
+
a
11
β
(a
02
+ a
20
) −
b
11
β
(b
02
+ b
20
)
_
.
For an explanation of where this mysterious quantity comes from, see Section
3.4 of the text of Perko [8]. Speciﬁcally, read the discussion of Poincaré maps (a
test for the stability of periodic solutions).
The Lyapunov number provides a quick way of testing whether the sys
tem (5.11) actually experiences an AndronovHopf bifurcation at µ = 0.
Theorem 5.3.5. (Hopf Bifurcation Theorem). If σ ,= 0, then the system (5.11)
experiences an AndronovHopf bifurcation at µ = 0. If σ < 0, a unique, stable limit
bi furcati ons 159
cycle bifurcates from the origin as µ increases from negative to positive. For small,
positive µ, the amplitude of the limit cycle solution is approximately proportional
to
√
µ. Similarly, if σ > 0, then a unique unstable limit cycle bifurcates from the
origin as µ decreases from positive to negative.
Example 5.3.6. In our previous example, we found that the system (5.10)
x
/
= −y + x(µ −x
2
−y
2
)
y
/
= x + y(µ −x
2
−y
2
)
suffers an AndronovHopf bifurcation at µ = 0. In our above notation, β = 1,
p(x, y) = −x
3
−xy
2
and q(x, y) = −x
2
y −y
3
. Notice that there are no “quadratic”
terms—i.e., terms in which the exponents of x and y sum to 2. This implies that
a
20
, a
11
, a
02
, b
20
, b
11
, and b
02
are all 0. For the “cubic” terms, we have
a
30
= −1, a
12
= −1, a
21
= 0, a
03
= 0,
and
b
30
= 0, b
12
= 0, b
21
= −1, b
03
= −1.
The Lyapunov number is
σ =
3π
2
[3(−1 −1) + (−1 −1)] = −12π < 0,
and it follows that a supercritical AndronovHopf bifurcation occurs at µ = 0.
It is possible to state a stronger version of the Hopf Bifurcation Theorem which
allows us to estimate the period and amplitude of the limit cycles that are born via
an AndronovHopf bifurcation. Suppose that a system x = f (x; µ) experiences
an AndronovHopf bifurcation at some critical parameter value µ = µ
c
. Let
λ(µ) = α(µ) + iω(µ) denote the eigenvalues of a complex conjugate pair of
eigenvalues whose real parts change sign at µ = µ
c
. At the bifurcation point,
ω(µ
c
) ,= 0 and α(µ
c
) = 0. Moreover, the eigenvalues satisfy a transversality
condition α
/
(µ
c
) ,= 0, which basically just states that the eigenvalues really are
moving from negative to positive real part (or viceversa) as µ passes µ
c
. Let us
consider the supercritical case in which α
/
(µ
c
) > 0, so that the real parts of the
pair of eigenvalues are changing from negative to positive. The Hopf Bifurcation
Theorem states that for µ slightly larger than µ
c
, there exists a family of stable
160 andronov hopf bi furcati ons
limit cycle solutions whose amplitudes are roughly proportional to
√
µ −µ
c
.
Moreover, if µ −µ
c
is reasonably small, the period of the limit cycle solutions is
approximately 2π/ω(µ
c
). For details, refer to the text of Hassard et al. [4].
As an illustration, suppose that the Jacobian matrix associated with a system
of odes has eigenvalues (µ − 3) ± 2µi. Then α(µ) = µ − 3 and ω(µ) = 2µ.
At µ = µ
c
= 3, we have α(3) = 0 and ω(3) = 6 ,= 0. The transversality
condition α
/
(3) = 1 ,= 0 is also satisﬁed, and we conclude that an Andronov
Hopf bifurcation has occurred at µ = 3. The bifurcation would be supercritical
because α
/
(3) > 0. This means that for µ values slightly larger than 3, there
exists a family of stable periodic solutions. The periods of these limit cycles are
approximately 2π/ω(3) = π/3, and their amplitudes are roughly proportional
to
_
µ −3. The bifurcation diagram for such a system would look essentially
identical to the one in Figure 5.13, except that the vertex of the paraboloidlike
surface would be shifted from µ = 0 to µ = 3.
Here, we will not attempt to provide a comprehensive list of the various
types of bifurcations that can occur. Clearly, there are many ways that the
qualitative behavior of a phase portrait could be suddenly and dramatically
altered as a parameter µ is varied. Systems with more than one parameter offer
an even greater degree of ﬂexibility. Students interested in learning more about
bifurcation theory are encouraged to take an advanced course in differential
equations and dynamical systems.
Exercises
1. For each of the following odes, µ denotes a real parameter. In each case,
identify the equilibria of the ode and determine which ones are stable and
unstable. Then, produce a bifurcation diagram.
(a)
dx
dt
= (µ −1)x + x
2
(b)
dx
dt
= µ −2 +3x
2
(c)
dx
dt
= µ
2
−x
2
(d)
dx
dt
= µx +9x
3
(e)
dx
dt
= 4x(µ −e
x
).
bi furcati ons 161
2. For each of the following odes, µ denotes a real parameter. In each case, ﬁnd
all bifurcation points and, if possible, classify them as one of the types we
discussed in this Chapter. Then, produce a bifurcation diagram.
(a)
dx
dt
= (1 +µ)(x
2
−µx)
(b)
dx
dt
= µ −x
2
+4x
4
(c)
dx
dt
= µ
2
−x
4
(d)
dx
dt
= (−x
2
+µ
4
)(x
2
−µ
2
).
3. Sketch the bifurcation diagram for the equation
dx
dt
= (µ −x)(µ + x
2
) = µ
2
−µx +µx
2
−x
3
,
where µ is a real parameter. If you do this correctly, you will discover two
different bifurcation points: one transcritical, and one pitchfork.
4. Let µ be a real parameter.
(a) Find all bifurcation points for the equation x
/
= x
2
−µx −2µ
2
, and draw
a bifurcation diagram.
(b) Suppose we modify the ode in Part (a), multiplying the righthand side
by x to obtain the equation x
/
= x(x
2
− µx − 2µ
2
). Sketch the new
bifurcation diagram and compare it to the bifurcation diagram from Part
(a).
(c) Although the system in Part (b) has no bifurcation points, its “bifurcation
diagram” still conveys useful information. In fact, the diagram completely
characterizes how the system behaves depending upon the parameter µ
and the initial condition x
0
. If µ ,= 0, note that there are two stable equi
libria. For µ ,= 0, ﬁnd the basin of attraction of each stable equilibrium. The
basin of attraction for an equilibrium x
∗
is the set of all initial conditions
x
0
for which x(t) → x
∗
as t →∞.
Remark. (i) Notice that the basins of attraction of the stable ﬁxed points are
“separated” by an unstable ﬁxed point. In this respect, unstable equilibria can
be very important, as they help dictate which stable state (if any) our system
will converge to as t →∞. (ii) Suppose that f is continuously differentiable
162 andronov hopf bi furcati ons
and that the ﬁrstorder ode x
/
= f (x; µ) has multiple stable equilibria. Then
there is always an unstable equilibrium between each pair of stable equilibria.
Any solid curves in the bifurcation diagram must be separated by a dashed
curve somewhere in between.
5. A “quadfurcation”: When we discussed saddlenode bifurcations, we gave
an example of an ode that has two equilibria for µ < 0 and no equilibria for
µ > 0. Create an example of an ode with a single parameter µ which has no
equilibria for µ < 0 but four equilibria for µ > 0. Identify which equilibria are
stable/unstable.
6. A “ﬁvepronged pitchfork”: When we discussed pitchfork bifurcations, we
gave an example of an ode that has one equilibrium for µ < 0 and three
equilibria for µ > 0. Create an example of an ode with a single parameter
µ which has one equilibrium for µ < 0 but ﬁve equilibria for µ > 0. Identify
which equilibria are stable/unstable.
7. Consider the system
dx
dt
= −8y + x(µ −x
2
−y
2
)
dy
dt
= 8x + y(µ −x
2
−y
2
),
where µ is a parameter.
(a) Show that, regardless of µ, the origin is the only equilibrium.
(b) Show that a Hopf bifurcation occurs when µ = 0.
(c) Give a qualitative description of the periodic solutions which are created
by this bifurcation. Estimate their period, assuming that µ is small. For
small µ, what can you say about the amplitude of the periodic solutions?
Are these solutions stable or unstable?
8. Subcritical Hopf Bifurcations. Subcritical Hopf bifurcations can be dangerous
in engineering applications because when an equilibrium loses stability as a
parameter µ varies, solutions can suddenly jump to a far distant stable limit
cycle. Consider, for example, the planar system in polar coordinates
dr
dt
= µr + r
3
−r
5
dθ
dt
= 1.
where µ is a bifurcation parameter.
bi furcati ons 163
(a) Note that r = 0 is an equilibrium. Determine the range of µ values for
which this equilibrium is stable/unstable.
(b) If we write the r equation as r
/
= r(µ+r
2
−r
4
), we see that other equilibria
must satisfy µ = r
4
−r
2
. Carefully sketch this curve in the µrplane. Hint:
You may ﬁnd it easier to ﬁrst sketch µ versus r, and then ﬂip your axes to
create a plot of r versus µ.
(c) Show that the r equation has two positive equilibria if −
1
4
< µ < 0 but
only one positive equilibrium if µ ≥ 0.
(d) Without attempting to ﬁnd formulas for the positive equilibria, show that
if µ > 0 then the positive equilibrium is stable.
(e) For −
1
4
< µ < 0, show that the smaller of the two positive equilibria is
unstable.
(f) Discussion (No work required on your part!) Here is why this sort of
bifurcation can be devastating. Suppose µ < −1/4 and our system is in
equilibrium r = 0. Now start gradually increasing µ. When we pass µ =
−1/4, two limit cycles are born, corresponding to the two different positive
r values which satisfy the equation µ = r
4
−r
2
. The limit cycle with larger
amplitude (larger r) is stable, and the other one is unstable. Meanwhile,
our equilibrium r = 0 is still stable. However, when we reach µ = 0, the
unstable limit cycle is destroyed via a subcritical Hopf bifurcation, but the
larger amplitude stable limit cycle still exists. As soon as the unstable limit
cycle is destroyed (at µ = 0), the equilibrium r = 0 also loses its stability.
Consequently, once µ > 0, any small amount of “noise” would throw us
off the equilibrium and our solution would (rather dramatically) leap to
a largeamplitude periodic solution. Physically, you can imagine a very
wellbehaved system which suddenly begins to oscillate violently as a
parameter is increased. Subcritical Hopf bifurcations have been observed
in mathematical models of aeroelastic ﬂutter (vibrations in airplane wings),
and a host of other physical/biological scenarios.
This system also provides a nice example of bistability and hysteresis, which
can be explained as follows. Again, suppose µ < −1/4 and we are at the
stable equilibrium. As we increase µ past 0, our equilibrium suddenly re
pels nearby trajectories and suddenly forces us toward the largeamplitude
stable, periodic solution. Now suppose we start to decrease µ again. This
time, we have a nice, stable limit cycle until we reach µ = −1/4, when we
164 andronov hopf bi furcati ons
suddenly transition back to a stable equilibrium. Notice that the sudden
transitions between two different stable states occurred at two different µ
values. This phenomenon is known as hysteresis, and the fact that there
are two different stable states for −1/4 < µ < 0 is known as bistability.
9. Selfoscillations in glycolysis: For particular parameter choices, a reduced
form of the famous Selkov model takes the form
dx
dt
= −x +
1
10
y + x
2
y
dy
dt
= µ −
1
10
y −x
2
y.
(a) Show that
(x, y) =
_
µ,
10µ
10µ
2
+1
_
is an equilibrium.
(b) Show that this equilibrium is a stable node if µ = 0.
(c) Show that this equilibrium is an unstable focus if µ =
_
1
2
.
(d) Show that this equilibrium is a stable focus if µ = 1.
(e) Since the eigenvalues of the Jacobian matrix depend continuously on µ, the
above observations suggest that at least two AndronovHopf bifurcations
have occurred: one between µ = 0 and µ =
_
1
2
and one between µ =
_
1
2
and µ = 1. Find the µ values at which the AndronovHopf bifurcations
occur. Remark: The ﬁrst bifurcation creates a family of stable, periodic
solutions. The periodic solutions disappear after the second bifurcation,
when the stable focus is born.
10. The purpose of this exercise is to create a bifurcation diagram for the system
dx
dt
= µx(1 −x) −xy
dy
dt
= x −y,
where µ > −1 is a real parameter.
(a) The origin is an equilibrium independent of the choice of µ. Find the
other equilibrium of the system.
(b) By linearizing the system at the origin, determine the ranges of µ for
which the origin is stable/unstable.
bi furcati ons 165
(c) Let (x
1
, y
1
) denote the equilibrium you found in Part (a). Compute the
Jacobian matrix J f (x
1
, y
1
) and use its determinant to ﬁnd the range of µ
for which the equilibrium is a saddle.
(d) When the determinant of J f (x
1
, y
1
) is positive, the origin is either a focus
or a node. However, for the purposes of creating a bifurcation diagram,
we need only determine whether the equilibrium is stable or unstable,
and this is easily accomplished by inspecting the trace and determinant of
J f (x
1
, y
1
). Show that the trace is negative for all µ > −1, and conclude
that if the determinant is positive, then (x
1
, y
1
) is a stable node.
(e) Since the x and y coordinates of the equilibria (0, 0) and (x
1
, y
1
) happen to
be equal, we need not create a threedimensional bifurcation diagram by
plotting µ versus both x and y. Because no information is lost if we neglect
the y variable, sketch a bifurcation diagram of µ versus x. What are the
bifurcation values of µ? Can you classify which type(s) of bifurcations
occur?
CHAPTER 6
Introduction to Delay Diﬀerential Equations
I
n this Chapter, we turn our attention to delay differential equations (ddes), a
major departure from the ordinary differential equations that were considered
up to now. A basic reference for this material is the text of Bellman and Cooke [2].
To understand why ddes are of mathematical interest, let us examine the simplest
population growth model, which was originally proposed by Malthus. The major
underlying assumption of the Malthus model is that the rate of change of
population is proportional to the population itself. Mathematically, let P(t)
denote the population at time t. Then the population growth model is given by
dP
dt
= kP,
where k is a positive constant. The solution of this ode is P(t) = P(0)e
kt
, which
predicts exponential population growth as t increases. However, due to the time
lag between conception and birth, it may be more realistic to assume that the
instantaneous rate of change of population growth is actually dependent upon
the population at some ﬁxed amount of time τ in the past. This would suggest
that we adjust the above model to read
d
dt
P(t) = kP(t −τ). (6.1)
Observe that the rate of change of P at time t is affected by the value of P at time
t −τ.
166
i ntroducti on to delay di fferenti al equati ons 167
Deﬁnition 6.0.7. Suppose x and t are dependent and independent variables,
respectively, and let τ be a positive constant. Any equation of the form
F(x(t), x
/
(t), x(t −τ), x
/
(t −τ), t) = 0
is called a ﬁrstorder dde with a single, constant delay. If the equation does not
incorporate x
/
(t −τ), the dde is called retarded. If the equation does incorporate
x
/
(t −τ), the dde is called neutral.
In the above deﬁnition, “ﬁrstorder” refers to the fact that ﬁrst derivatives are
the highestorder derivatives that appear in the equation. The words “single,
constant delay” refer to the fact the equation only makes reference to the present
time, t, and one ﬁxed time in the past, t −τ.
Example 6.0.8. The equation
dx
dt
= x
2
−(x −3) + x(t −2)
is a retarded ﬁrstorder dde with a single constant delay τ = 2. On the right
hand side, it is understood that x
2
means x(t)
2
and that (x −3) means x(t) −3.
The equation
x
/
(t) = x
2
−(x −3) + x(t −2) + x(t −4)
is a retarded, ﬁrstorder dde with two constant delays: τ
1
= 2 and τ
2
= 4. The
equation x
/
(t) = x(t/2) is a dde with a variable time delay. Note that the rate of
change of x when t = 1 is inﬂuenced by the value of x at time t = 1/2, whereas
the rate of change of x when t = 6 is inﬂuenced by the value of x at time t = 3.
Finally, the equation
x
/
(t −8) + x(t −8) + x(t) + t = 0
is a neutral dde with a single constant delay τ = 8.
This Chapter will focus exclusively on retarded, ﬁrstorder ddes with a single,
constant time delay τ.
168 i ni ti al value problems
6.1. Initial Value Problems
If we wish to solve a dde such as x
/
(t) = x(t −1), how would we specify initial
data? Note that the solution at time t −1 inﬂuences the rate of change at time t.
So, for example, in order to know the rate of change of x for t ∈ [0, 1], we would
need to know the value of x(t) for t ∈ [−1, 0]. That is, before we can generate
the solution of this dde on the interval 0 ≤ t ≤ 1, we must require that initial
data be provided as a function on the entire interval −1 ≤ x ≤ 0. More generally,
in order to solve a retarded dde with constant time delay τ, we must specify an
initial function φ(t) on the interval [−τ, 0].
Method of Steps. We illustrate one method for solving a dde via an example.
Consider the system
x
/
(t) = x(t −1), if t > 0
x(t) = φ(t) = 1, if −1 ≤ t ≤ 0.
Here, we have speciﬁed the “initial function” φ(t) in a closed interval of width 1
since our time delay is τ = 1. To solve the dde, note that by the Fundamental
Theorem of Calculus we have
x(t) = x(0) +
_
t
0
x
/
(s) ds.
In this case, x
/
(s) = x(s −1), and if 0 ≤ s ≤ 1, then we have −1 ≤ (s −1) ≤ 0.
So for 0 ≤ s ≤ 1, it follows that x(s −1) = φ(s −1) = 1, and the above equation
reduces to
x(t) = x(0) +
_
t
0
1 ds = 1 + t, (0 ≤ t ≤ 1).
Now that we have obtained the solution x(t) = 1 + t on the interval 0 ≤ t ≤ 1,
we repeat the procedure to solve for x(t) on the next interval, 1 ≤ t ≤ 2. For
t ∈ [1, 2], we calculate
x(t) = x(1) +
_
t
1
x
/
(s) ds = x(1) +
_
t
1
x(s −1) ds.
i ntroducti on to delay di fferenti al equati ons 169
Making a substitution to replace (s −1) with s, the latter integral becomes
x(t) = x(1) +
_
t−1
0
1 + s ds = 2 + (t −1) +
(t −1)
2
2
.
Therefore,
x(t) =
t
2
2
+
3
2
if 1 ≤ t ≤ 2.
Continuing in this fashion, you can calculate that for 2 ≤ t ≤ 3,
x(t) = x(2) +
_
t
2
x(s −1) ds = x(2) +
_
t
2
(s −1)
2
2
+
3
2
ds, etc.. . .
This technique of extending the solution one interval at a time is called the
method of steps. In this particular example, we can write the solution for t ≥ 0 by
introducing the notation ¸t to denote the largest integer which is smaller than
or equal to t. By induction, one may argue that the solution of our dde is given
by
x(t) =
¸t+1
∑
n=0
[t −(n −1)]
n
n!
for t ≥ 0.
6.2. Solving ConstantCoeﬃcient Delay Diﬀerential Equations
Recall the variation of parameters formula (2.18), which says that the ﬁrstorder,
linear ode x
/
= Ax +b(t) has solution
x(t) = e
tA
x
0
+ e
tA
_
t
0
e
−sA
b(s) ds.
This formula is easily adapted to yield the solution of the general ﬁrstorder,
constantcoefﬁcient retarded dde
x
/
(t) = ax(t) + bx(t −τ) + f (t) if t > 0,
x(t) = φ(t) if t ∈ [−τ, 0].
(6.2)
170solvi ng constant coeffi ci ent delay di fferenti al equati ons
Here a and b are constants, τ > 0 is the time delay, and f (t) is a continuous
function. For 0 ≤ t ≤ τ, the solution to (6.2) is given by
x(t) = e
at
x(0) +e
at
_
t
0
e
−as
[bx(s −τ) + f (s)] ds
= e
at
φ(0) +e
at
_
t
0
e
−as
[bφ(s −τ) + f (s)] ds.
Notice that this formula allows us to express x(t) in terms of the known functions
φ and f , at least on the interval [0, τ]. The method of steps can then be used to
extend the solution to the interval [τ, 2τ] and so on.
We make two remarks regarding solutions of ddes. First, notice that solutions
need not be differentiable at the endpoints of consecutive intervals. In our
introductory example of the method of steps, we had x(t) = 1 for −1 ≤ t ≤ 0,
but x(t) = 1 + t for 0 ≤ t ≤ 1. Although the solution is continuous at t = 0, it is
not differentiable there. Second, notice that the very construction of the above
solution implies that constantcoefﬁcient ddes have a unique solution.
Provided that the functions φ and f are continuous, we know that the sys
tem (6.2) has a unique solution for all t ≥ 0, and the method of steps can be
used to construct that solution. Under what conditions on φ will the solution be
continuously differentiable for all time t ≥ 0? First, suppose that φ is differentiable
on [−τ, 0]. Since x(t) = φ(t) on that interval, we also have x
/
(t) = φ
/
(t). If we
want the solution to be differentiable at t = 0, then we must insist that left hand
and right hand derivatives of x(t) match at t = 0. Let us introduce the notation
x(0
−
) = lim
t→0
−
x(t) and x(0
+
) = lim
t→0
+
x(t).
Note that x
/
(0
−
) = φ
/
(0
−
) since x
/
(t) = φ
/
(t) when t < 0. On the interval [0, τ],
the dde tells us that
x
/
(t) = ax(t) + bx(t −τ) + f (t) = ax(t) + bφ(t −τ) + f (t)
Taking the right hand limit, we ﬁnd that
x
/
(0
+
) = ax(0
+
) + bφ(−τ
+
) + f (0
+
).
i ntroducti on to delay di fferenti al equati ons 171
But since f is continuous, it follows that f (0
+
) = f (0), and moreover x(0) = φ(0).
Equating the left and right hand derivatives at t = 0, we obtain the condition
φ
/
(0) = aφ(0) + bφ(−τ) + f (0). (6.3)
Theorem 6.2.1. The solution of (6.2) is continuously differentiable for all t ≥ 0 if
and only if φ is differentiable at t = 0 and condition (6.3) is satisﬁed.
The condition (6.3) actually leads to some remarks about the “smoothness” of
solutions of ddes. For retarded ddes with constant delays, any initial disconti
nuities are smoothed out as we advance forward in time. Neutral ddes do not
have this luxury—if the solution is discontinuous at t = 0, then the solution will
also be discontinuous at the endpoint of each interval [nτ, (n + 1)τ] for each
positive integer n. The method of steps can still be applied to neutral ddes, but
using a computer to numerically approximate solutions of such ddes can be very
challenging.
Linear, constantcoefﬁcient ddes can also be solved using Laplace transform
methods. For details, see Bellman and Cooke [2].
6.3. Characteristic Equations
Recall from your introductory course in odes that for constantcoefﬁcient prob
lems such as x
//
+ 6x
/
+ 8x = 0, we expect exponential solutions of the form
x = e
λt
, where λ is a constant. Substituting this exponential function into the
ode leads to the equation
λ
2
e
λt
+6λe
λt
+8e
λt
= e
λt
(λ
2
+6λ +8) = 0.
Since the exponential factor could never be 0, it must be the case that λ
2
+6λ +
8 = 0. This is called the characteristic equation for this ode and, in this case, its
roots are λ = −2 and λ = −4. Consequently, e
−2t
and e
−4t
are solutions of the
ode, as is any linear combination C
1
e
−2t
+ C
2
e
−4t
.
Now consider the homogeneous, constantcoefﬁcient dde given by x
/
(t) =
ax(t) + bx(t −τ), where a and b are constants and τ is a positive, constant time
delay. As above, let us seek exponential solutions x = e
λt
. Then substitution
yields
λe
λt
= ae
λt
+ be
λ(t−τ)
= ae
λt
+ be
λt
e
−λτ
.
172 the hutchi nson wri ght equati on
Dividing by e
λt
, we obtain the characteristic equation
λ −a −be
−λτ
= 0.
Bad news: This is a transcendental equation for λ, and it is impossible to alge
braically solve for λ in terms of the constants a, b, and τ.
More bad news: Excluding the silly case b = 0 (in which the dde would actually
have been an ode), the characteristic equation has inﬁnitely many complexvalued
solutions.
Slightly better news: If we could ﬁnd all of the roots of the characteristic
equation, then we could write the general solution of the dde as an inﬁnite sum
of various exponential functions, provided that we can show that such series
converge.
Relative to constantcoefﬁcient odes, working with characteristic equations for
ddes is much more challenging. However, doing so can still be quite illuminating
when we try to get a feel for the qualitative behavior of solutions. The remainder
of this Chapter is dedicated to illustrating this via an example.
6.4. The HutchinsonWright Equation
In this section, we perform a qualitative analysis of the dynamical behavior of a
nonlinear dde. The following calculations appear in the text of Hassard et al. [4]
One of the most common models for population growth of a single species is
given by the logistic equation
dP
dt
= rP
_
1 −
P
K
_
,
where r and K are positive constants representing the population growth rate
and maximum sustainable population, respectively. This ode can be solved
analytically via separation of variables. Qualitatively, it is easy to see that
the ode has two equilibria: P = 0 (corresponding to extinction) and P = K
(corresponding to the maximum population that the environment can sustain).
The ﬁrst of these equilibria is unstable, and the second is stable.
i ntroducti on to delay di fferenti al equati ons 173
Now suppose that we incorporate a time delay into this equation, accounting
for the lag between conception and birth. Hutchinson modiﬁed the logistic
equation as
dP
dt
= rP(t)
_
1 −
P(t −τ)
K
_
, (6.4)
where τ > 0 is a positive, constant time delay. This equation is now known
as the HutchinsonWright equation. It is a nonlinear dde and cannot be solved
analytically.
Equilibria: We begin our qualitative analysis of (6.4) by noting that there are
still only two equilibria, P = 0 and P = K.
Stability: It is unclear how (or whether) we could extend our usual “eigenvalues
of the Jacobian” test for stability from odes to ddes like (6.4). The whole idea of
stability analysis is to analyze how a system in equilibrium would respond to a
small perturbation. First consider the equilibrium P = 0, and suppose that we
seek a solution that is “close” to equilibrium: P(t) = 0 +y(t) where is a small,
positive number. We linearize the HutchinsonWright equation by substituting
that expression into the dde:
dy
dt
= ry(t)
_
1 −
y(t −τ)
K
_
.
Dividing by and expanding the righthand side, we obtain the dde
dy
dt
= ry(t) −
r
K
y(t)y(t −τ).
Since is small by assumption, the linearization of the dde is obtained by
neglecting the latter term. The result is an ode, namely
dy
dt
= ry, which is the
linearization of the HutchinsonWright equation at the equilibrium P = 0. This
ode is easily solved, and we immediately see that the origin is an unstable
equilibrium since r is a positive constant.
The linearization at the other equilibrium P = K is a bit more difﬁcult. As
before, we wish to examine a small perturbation from the equilibrium. Let
P(t) = K +y(t), where is small and positive. Substituting that expression into
174 the hutchi nson wri ght equati on
equation (6.4),
d
dt
(K + y(t)) = r[K + y(t)]
_
1 −
K + y(t −τ)
K
_
.
By algebra,
dy
dt
= −r[K + y(t)]
_
y(t −τ)
K
_
.
Expanding the righthand side and dividing by reveals that
dy
dt
= −ry(t −τ) −
r
K
y(t −τ)y(t).
Since is assumed to be small, we may obtain the linearized system by neglecting
the rightmost term:
y
/
(t) = −ry(t −τ).
This ﬁrstorder constantcoefﬁcient dde is the linearization of (6.4) at the equi
librium P = K, and can actually be solved via the method of steps. As we shall
see, it turns out that P = K is a stable equilibrium if rτ < π/2 and is unstable if
rτ > π/2. This raises the following
Observation. Recall that P = K was always a stable equilibrium for the logistic
ode that we discussed at the beginning of this section. The inclusion of a time
delay τ in the model has a destabilizing effect on the system, because P = K
becomes an unstable equilibrium if τ is appropriately large.
Question: What exactly happens when rτ = π/2 that causes the equilibrium
P = K to suddenly lose stability? It must be some sort of bifurcation, but what
type?
To answer this question and further analyze the stability of P = K, it will be
convenient to rescale the variables in the HutchinsonWright equation. First,
we will rescale the time variable in such a way that the time delay is equal
to 1: let s = t/τ be our new time variable. Next, we will introduce another
rescaling which both (i) eliminates the need for the parameter K and (ii) moves
the equilibrium from P = K to P = 0; namely,
x(s) =
P(τs) −K
K
.
i ntroducti on to delay di fferenti al equati ons 175
With our new independent variable s and dependent variable x, the dde (6.4)
becomes
d
ds
x(s) = −(rτ)x(s −1)[1 + x(s)].
Finally, based upon the earlier claim that a bifurcation occurs when rτ = π/2,
it is convenient to “recenter” the bifurcation so that it will occur at 0. To
accomplish this, introduce a new parameter µ = rτ −π/2 so that the dde takes
the form
x
/
(s) = −
_
µ +
π
2
_
x(s −1)[1 + x(s)]. (6.5)
Note that x = 0 is an equilibrium of this system, and it corresponds to the
equilibrium P = K of the original dde (prior to rescaling).
To see that a bifurcation really does occur when µ = 0, causing the equilibrium
x = 0 to lose stability, examine the linearization of (6.5) at that equilibrium:
x
/
(s) = −
_
µ +
π
2
_
x(s −1).
The characteristic equation is obtained by substituting x = e
λs
:
λe
λs
= −
_
µ +
π
2
_
e
λ(s−1)
= −
_
µ +
π
2
_
e
λs
e
−λ
.
Dividing by e
λs
, the characteristic equation is
λ +
_
µ +
π
2
_
e
−λ
= 0. (6.6)
This is a transcendental equation, and has inﬁnitely many complexvalued roots.
To test whether any AndronovHopf bifurcations occur as the parameter µ is
varied, we wish to determine conditions under which the characteristic equation
has pure imaginary roots λ = ±iω, where ω > 0 is real. After all, we know that
AndronovHopf bifurcations occur when a complex conjugate pair of eigenvalues
switches from negative to positive real part, or viceversa. Substituting λ = iω
into the characteristic equation reveals that
iω +
_
µ +
π
2
_
(cos ω −i sin ω) = 0, (6.7)
where we have used Euler’s identity to write e
−iω
= cos ω −i sin ω. Equa
tion (6.7) can only be satisﬁed if both the real and imaginary parts of the
176 the hutchi nson wri ght equati on
expressions on the lefthand side are both equal to 0. Therefore,
_
µ +
π
2
_
cos ω = 0 and ω −
_
µ +
π
2
_
sin ω = 0.
Since we originally claimed that a bifurcation occurs when µ = 0, focus attention
on that particular parameter choice. The two equations become
π
2
cos ω = 0 and ω =
π
2
sin ω.
Substituting the second of these into the ﬁrst,
π
2
cos
_
π
2
sin ω
_
= 0.
This requires sin ω = ±1, and we already assumed that ω > 0. The only
possible choices for ω are odd positive integer multiples of π/2. That is
ω = (2n +1)
π
2
n = 0, 1, 2, . . . .
In particular, we have found that for µ = 0, the characteristic equation has pure
imaginary roots ±
π
2
i. Note that due to the periodicity of the sine and cosine
functions, we have
e
(π/2)i
= e
(5π/2)i
= e
(9π/2)i
= · · ·
and
e
(−π/2)i
= e
(3π/2)i
= e
(7π/2)i
= · · ·
Hence, we need only direct our attention towards the speciﬁc roots ±
π
2
.
The fact that the characteristic equation has pure imaginary roots when µ = 0
suggests that an AndronovHopf bifurcation may occur at that critical parameter
value. According to the remarks about the Hopf Bifurcation Theorem at the
end of the preceding Chapter, it remains only to verify that the transversality
condition holds at µ = 0; i.e., that the signs of the real parts of a pair of complex
conjugate eigenvalue really does change from negative to positive (or viceversa)
when µ passes through 0. Indeed, we will show that the real part of
dλ
dµ
is
nonzero when µ = 0. To compute
dλ
dµ
, we use implicit differentiation of the
i ntroducti on to delay di fferenti al equati ons 177
characteristic equation (6.6)
λ +
π
2
e
−λ
+µe
−λ
= 0.
Differentiating both sides with respect to µ,
dλ
dµ
+
_
π
2
_
_
−
dλ
dµ
_
e
−λ
+ e
−λ
+ µ
_
−
dλ
dµ
_
e
−λ
= 0.
Algebraically solving for
dλ
dµ
reveals that
dλ
dµ
=
−e
−λ
1 −
_
π
2
+µ
_
e
−λ
.
Recall that when µ = 0, we know that there is a pair of pure imaginary eigenvalue
λ = ±
π
2
i. Substituting these values into the last equation,
dλ
dµ
¸
¸
¸
¸
µ=0
= −
e
−(π/2)i
1 −
_
π
2
_
e
−(π/2)i
.
By Euler’s identity, e
−(π/2)i
= cos(−π/2) +i sin(−π/2) = −i. Thus,
dλ
dµ
¸
¸
¸
¸
µ=0
=
i
1 +
π
2
i
.
Multiplying and dividing the denominator of this last expression by its conjugate
yields
dλ
dµ
¸
¸
¸
¸
µ=0
=
π
2
+i
1 +
1
4
π
2
.
The real part of
dλ
dµ
is
π
2
1 +
1
4
π
2
> 0.
Consequently, as µ increases from negative to positive, the real parts of our
pair of eigenvalues λ is also changing from negative to positive. A supercritical
AndronovHopf bifurcation occurs, and creates stable, periodic solutions for µ > 0.
According to our remarks at the end of the preceding chapter, for µ slightly
larger than 0, the amplitude of these periodic solutions should be (roughly)
proportional to
√
µ. The period of the oscillations is estimated by dividing 2π
178 the hutchi nson wri ght equati on
by the imaginary part of the eigenvalues at the bifurcation point µ = 0. In this
case, the period would be approximately 2π/(π/2) = 4. For our rescaled
HutchinsonWright equation, there are known approximations of the periodic
solutions that are created when µ > 0. Speciﬁcally, if µ is small and positive,
then
x(s) =
_
40
3π −2
√
µ cos
_
π
2
s
_
is an approximate solution, and the error in this approximation is roughly
proportional to µ.
Observation. The HutchinsonWright equation predicts that the equilibrium
population P = K will lose stability if the growth rate r and/or the time delay
τ are too large, resulting in oscillations in population. The fact that a ﬁrstorder
dde can produce oscillations is noteworthy, because ﬁrstorder autonomous odes
cannot have [nonconstant] periodic solutions.
To see why, consider the autonomous, ﬁrstorder equation
dy
dt
= f (y) and
suppose that there is a periodic solution with period τ. Then y(t + τ) = y(t) for
all time t. Multiply both sides of the ode by
dy
dt
to get
_
dy
dt
_
2
= f (y)
dy
dt
.
Now integrate over one period:
_
t+τ
t
_
dy
dt
_
2
dt =
_
t+τ
t
f (y)
dy
dt
dt =
_
y(t+τ)
y(t)
f (y) dy.
The rightmost integral is zero because y(t) = y(t + τ). On the other hand, the
leftmost integral would have to be positive unless y were a constant function. It
follows that ﬁrstorder autonomous odes cannot have periodic solutions. In this
respect (and many others), ddes can exhibit far richer dynamical behavior than
odes.
Exercises
1. Check that y(t) = sin t (t ≥ 0) is a solution of the initial value problem
y
/
(t) = −y
_
t −
π
2
_
(t ≥ 0),
i ntroducti on to delay di fferenti al equati ons 179
y(t) = φ(t) = sin t (t ≤ 0).
2. Let ¸t denote the largest integer which is less than or equal to t. Use induction
to show that the solution of the initial value problem
y
/
(t) = Ay(t −B) (t ≥ 0),
y(t) = φ(t) = C (t ≤ 0)
is given by
y(t) = C
¸
t
B
+1
∑
n=0
A
n
[t −(n −1)B]
n
n!
for t ≥ 0. Here, A, B, and C are constants.
3. Show that the characteristic equation for the dde x
/
(t) = x(t −1) has exactly
one real root.
CHAPTER 7
Introduction to Diﬀerence Equations
T
his Chapter concerns the dynamical behavior of systems in which time
can be treated as a discrete quantity as opposed to a continuous one. For
example, some mathematical models of the onset of cardiac arrhythmias are
discrete, due to the discrete nature of the heartbeat. A more standard example
involves population models for species without overlap between successive
generations. If P
n
denotes the population of the nth generation, is there a
functional relationship P
n+1
= f (P
n
) which would allow us to predict the
population of the next generation? Below, we will learn techniques for analytical
and qualitative analysis of such systems. Good references for this material
include the texts of Elaydi [3] and Strogatz [11].
7.1. Basic Notions
For the discrete systems that we shall consider, time t is no longer a continuous
variable as in the case of odes. Instead, we will typically use a nonnegative
integer n to index our discrete time variable. If x is a dependent variable, we will
use subscripts x
n
instead of writing x(n) to represent the value of x at time n.
Example 7.1.1. An example of a discrete system is given by x
n+1
= x
2
n
. If we start
with an initial condition x
0
∈ R, then we may recursively determine the values
of all values in the sequence {x
n
}
∞
n=0
. If x
0
= 1/2, then x
1
= 1/4, x
2
= 1/16,
and so on.
Deﬁnition 7.1.2. A system of the form x
n
= f (x
n−1
, x
n−2
, . . . x
n−k
) is an example
of a kthorder difference equation. Such systems are sometimes called kdimensional
mappings.
180
i ntroducti on to di fference equati ons 181
The solution of a kth order difference equation is simply the sequence {x
n
}.
Notice that a kthorder difference equation recursively generates its iterates, and
x
n
is affected by x
n−1
, x
n−2
, . . . x
n−k
. In particular, k initial conditions would be
required in order to start the iterative process of solving the equation.
Example 7.1.3. The famous Fibonacci sequence is generated recursively by the
secondorder difference equation x
n+1
= x
n
+x
n−1
, with initial conditions x
0
= 1
and x
1
= 1. The next iterate is generated by summing the previous two iterates.
Thus x
2
through x
7
are given by 2, 3, 5, 8, 13, and 21.
Example 7.1.4. Wellposedness is generally not a major issue for difference
equations, because a kthorder difference equation with k initial conditions will
always generate a unique sequence of iterates, provided that f is wellbehaved.
However, if there are restrictions on the domain of f , some difﬁculties can arise.
Consider the ﬁrstorder equation x
n+1
= ln(x
n
) with initial condition x
0
= e.
Then x
1
= 1, x
2
= 0, and x
n
is undeﬁned for n ≥ 3.
Closed formulas. Above we listed the ﬁrst few iterates in the solution of x
n+1
=
x
2
n
with the initial condition x
0
= 1/2. Based upon the pattern exhibited by these
iterates, we are led to conjecture that
x
n
=
1
2
2
n
,
which can, indeed, be proved by straightforward induction on n. This formula
for x
n
is ideal in that it provides an exact formula for all of the iterates in the
solution of the initial value problem. Such formulas are called closed formulas
for the solution of the difference equation. Producing a closed formula for
the solution of a difference equation is usually too much to hope for, but for
constantcoefﬁcient systems, closed formulas are readily available.
7.2. Linear, ConstantCoeﬃcient Diﬀerence Equations
Recall from your course in basic differential equations that, for constantcoefﬁcient
odes, we typically seek exponential solutions of the form e
λt
, where λ is a con
stant that must be solved for. Substituting exponential functions into a linear,
homogeneous, constantcoefﬁcient ode yields a polynomial equation involving
λ: the socalled characteristic equation. By ﬁnding the roots of the characteristic
182 li near, constant coeffi ci ent di fference equati ons
polynomial, one may then build the general solution of the ode by taking linear
combinations of exponential functions. The same idea can be used to solve
constantcoefﬁcient difference equations such as the Fibonacci equation above.
The only difference is that instead of seeking exponential solutions, one seeks
power function solutions of the form x
n
= λ
n
. The “characteristic equation
approach” to solving difference equations is developed in the exercises at the
end of this chapter. Now, we shall introduce a more elegant method which
recycles all of the techniques we learned when solving constantcoefﬁcient sys
tems of odes. First, we remark that kthorder difference equations, like kth order
odes, can always be written as systems of k ﬁrstorder equations. The idea is
to introduce new variables to represent any iterate other than the immediately
preceding iterate. Let us clarify this vague remark via an example: consider the
secondorder difference equation x
n+1
= f (x
n
, x
n−1
). If we replace x
n−1
with a
new variable y
n
, then note that y
n+1
= x
n
. Thus, the secondorder equation we
started with may be written as a system
x
n+1
= f (x
n
, y
n
) and y
n+1
= x
n
.
The vector (x
n+1
, y
n+1
) of iterates at the (n +1)st time step is expressed in terms
of the vector (x
n
, y
n
) of iterates at the nth time step—a system of two ﬁrstorder
equations. Similarly, the thirdorder equation
x
n+1
= f (x
n
, x
n−1
, x
n−2
)
can be written as a system by introducing y
n
= x
n−1
and z
n
= x
n−2
. It follows
that y
n+1
= x
n
and z
n+1
= x
n−1
= y
n
. Thus, our thirdorder equation can be
written as a system of three ﬁrstorder difference equations:
x
n+1
= f (x
n
, y
n
, z
n
)
y
n+1
= x
n
z
n+1
= y
n
.
In this Section, we will learn to solve kthorder, homogeneous, constantcoefﬁcient
difference equations; i.e., equations of the form
x
n+1
= a
0
x
n
+ a
1
x
n−1
+ · · · + a
k−1
x
n−k+1
i ntroducti on to di fference equati ons 183
where a
0
, a
1
, . . . a
k−1
are constants and are independent of n. Some brief remarks
about the special cases k = 1 and k = 2 will illuminate the solution process.
If k = 1, then our constantcoefﬁcient difference equation can be written as
x
n+1
= ax
n
, where a is a constant. Obtaining a closed formula for the solution
is easy, because each iterate is a constant multiple of the previous iterate. The
solution is x
n
= a
n
x
0
. Now suppose that k = 2. Based upon our above remarks,
any secondorder constantcoefﬁcient difference equation can be written as a
system
x
n+1
= a
11
x
n
+ a
12
y
n
y
n+1
= a
21
x
n
+ a
22
y
n
.
As with systems of odes, it is convenient to introduce matrix/vector notation.
Letting
A =
_
a
11
a
12
a
21
a
22
_
,
our system becomes
_
x
n+1
y
n+1
_
= A
_
x
n
y
n
_
.
Notice that each vector of iterates is updated by multiplying the previous vector
of iterates by the coefﬁcient matrix A. Thus,
_
x
n
y
n
_
= A
_
x
n−1
y
n−1
_
= A
2
_
x
n−2
y
n−2
_
= · · · = A
n
_
x
0
y
0
_
,
which means that the closed form solution of such a system is obtained by ﬁrst
computing powers A
n
of the coefﬁcient matrix, and then multiplying by the
vector of initial conditions.
Observation. If we can ﬁnd the appropriate canonical form for A, it is easy to
compute powers of A. For example,
If A is diagonalizable, then we may write A = PDP
−1
where P is an invertible
matrix and D is diagonal. In this case, A
n
= PD
n
P
−1
.
If A is not diagonalizable but has real eigenvalues, then we may write A =
S + N where S is diagonalizable, N is nilpotent, and SN = NS. If we then
write S = PDP
−1
where D is diagonal, then powers of S can be computed from
S
k
= PD
k
P
−1
. Since S and N commute, then we may use the binomial theorem
184 li near, constant coeffi ci ent di fference equati ons
to calculate
A
n
= (S + N)
n
=
n
∑
k=0
_
n
k
_
S
k
N
n−k
=
n
∑
k=0
_
n
k
_
PD
k
P
−1
N
n−k
.
The fact that N is nilpotent implies that many of the terms in this sum will likely
vanish, because large powers of N will always be 0. In principle, this sum could
be computed to obtain a closed formula for the solution, although doing so
would be tedious.
If A has complex conjugate eigenvalues, then ﬁnding the real canonical form
for A would facilitate computing large powers of A.
We now illustrate these ideas via examples.
Example 7.2.1. Solve the system
_
x
n+1
y
n+1
_
=
_
1 −1
2 4
_ _
x
n
y
n
_
where
_
x
0
y
0
_
=
_
3
−1
_
.
Solution: Let A denote the coefﬁcient matrix. The characteristic equation is
λ
2
−5λ +6 = 0 and, by factoring, we ﬁnd that the eigenvalues are λ = 2 and
λ = 3. You should check that
_
−1
1
_
and
_
1
−2
_
are eigenvectors for λ = 2 and λ = 3, respectively. Hence, we may write
A = PDP
−1
where
P =
_
−1 1
1 −2
_
, D =
_
2 0
0 3
_
, P
−1
=
_
−2 −1
−1 −1
_
.
The powers of A are given by
A
n
= PD
n
P
−1
= P
_
2
n
0
0 3
n
_
P
−1
=
_
2
n+1
−3
n
2
n
−3
n
−2
n+1
+2(3
n
) −2
n
+2(3
n
)
_
.
i ntroducti on to di fference equati ons 185
Multiplying by the vector of initial conditions, the solution of the initial value
problem is given by
_
x
n
y
n
_
= A
n
_
3
−1
_
=
_
3(2
n+1
) −3
n+1
−2
n
+3
n
−3(2
n+1
) +2(3
n+1
) +2
n
−2(3
n
)
_
=
_
5(2
n
) −2(3
n
)
(−5)(2
n
) +4(3
n
)
_
.
Notice that the solution involves powers of the eigenvalues. By contrast, we know
that solutions of systems of odes involve exponential functions with eigenvalues
appearing in the exponents.
Example 7.2.2. Solve the Fibonacci equation x
n+1
= x
n
+ x
n−1
with x
0
= x
1
= 1.
Solution: First, we write the equation as a system by introducing y
n
= x
n−1
.
The initial condition for y would be y
0
= x
−1
, which requires us to generate the
iterate that precedes x
0
, taking one step “backwards in time”. From Fibonacci’s
equation, it must be the case that x
−1
+ x
0
= x
1
, from which we conclude that
x
−1
= 0. The system now reads
_
x
n+1
y
n+1
_
=
_
1 1
1 0
_ _
x
n
y
n
_
and
_
x
0
y
0
_
=
_
1
0
_
.
As before, we let A denote the coefﬁcient matrix. The characteristic equation is
given by λ
2
−λ −1 = 0, which has roots
λ
±
=
1 ±
√
5
2
.
To ﬁnd an eigenvector for the positive eigenvalue λ
+
, note that
A −λ
+
I =
_
1 −λ
+
1
1 −λ
+
_
has reduced rowechelon form
_
1 −λ
+
0 0
_
.
186 li near, constant coeffi ci ent di fference equati ons
Eigenvectors v must satisfy v
1
= λ
+
v
2
, so if we treat v
2
as a free variable we see
that
_
λ
+
1
_
is an eigenvector for λ
+
. Recycling the same computations, it follows that
_
λ
−
1
_
is an eigenvector for λ
−
. As usual, we write A = PDP
−1
where
P =
_
λ
+
λ
−
1 1
_
, D =
_
λ
+
0
0 λ
−
_
,
and
P
−1
=
1
λ
+
−λ
−
_
1 −λ
−
−1 λ
+
_
.
Multiplying A
n
by the vector of initial conditions,
_
x
n
y
n
_
= A
n
_
x
0
y
0
_
= PD
n
P
−1
_
1
0
_
=
1
√
5
_
λ
+
λ
−
1 1
_ _
λ
n
+
0
0 λ
n
−
_ _
1 −λ
−
−1 λ
+
_ _
1
0
_
=
1
√
5
_
λ
n+1
+
−λ
n+1
−
λ
n
+
−λ
n
−
_
.
Notice that the second row of the solution vector is identical to the ﬁrst row, with
(n +1) replaced by n. This is not at all surprising if we recall that the variable y
n
was introduced to substitute for x
n−1
. Only the ﬁrst component of the solution
vector is important for our purposes, as it provides a closed formula for x
n
, the
solution of the Fibonacci equation:
x
n
=
1
√
5
_
_
_
1 +
√
5
2
_
n+1
−
_
1 −
√
5
2
_
n+1
_
_
.
i ntroducti on to di fference equati ons 187
You may ﬁnd it mildly surprising that the closed formula involves powers of the
irrational numbers λ
+
and λ
−
even though the Fibonacci sequence consists only
of positive integer values.
Example 7.2.3. Solve the initial value problem
_
x
n+1
y
n+1
_
=
_
2 1
0 2
_ _
x
n
y
n
_
and
_
x
0
y
0
_
=
_
2
1
_
.
Solution: The coefﬁcient matrix A has a repeated real eigenvalue λ = 2. By
itself, this is not enough to conclude that A is nondiagonalizable. However, if
you try to compute eigenvectors for A, you will ﬁnd that the eigenvalue only
has geometric multiplicity 1. Hence, A is not diagonalizable, and we must
write A = S + N where S is diagonalizable and N is nilpotent. This could be
accomplished by ﬁnding a generalized eigenvector for λ = 2; however, in this
case we may exploit the relatively simple form of A in order to write A = 2I + N
where I is the identity matrix and
N =
_
0 1
0 0
_
is nilpotent of order 2. Clearly 2I and N commute since I is the identity matrix.
Now, using the binomial theorem, we may compute powers of A:
A
n
= (2I + N)
n
=
n
∑
k=0
_
n
k
_
(2I)
n−k
N
k
=
n
∑
k=0
_
n
k
_
2
n−k
N
k
.
Since N is nilpotent of order 2, we know that N
2
(and all higher powers of N)
will be the zero matrix. Thus, only the ﬁrst two terms of this summation survive.
The relevant binomial coefﬁcients are (
n
0
) = 1 and (
n
1
) = n, and the summation
simpliﬁes to A
n
= 2
n
N
0
+ n2
n−1
N
1
. In other words,
A
n
=
_
2
n
n2
n−1
0 2
n
_
.
Finally, multiplying by the vector of initial conditions yields the closed formula
_
x
n
y
n
_
=
_
2
n+1
+ n2
n−1
2
n
_
.
188 li near, constant coeffi ci ent di fference equati ons
Example 7.2.4. Solve the initial value problem
x
n+1
= 3x
n
−18y
n
x
0
= 1
y
n+1
= 2x
n
−9y
n
y
0
= 1.
Solution: The characteristic equation associated with the coefﬁcient matrix
A =
_
3 −18
2 −9
_
is λ
2
+ 6λ + 9 = 0, and we see that λ = −3 is a repeated, real eigenvalue.
Eigenvectors v satisfy (A −λI)v = 0. The reduced rowechelon form of
A −λI = A +3I =
_
6 −18
2 −6
_
is
_
1 −3
0 0
_
.
It follows that
v =
_
3
1
_
is an eigenvector, and the eigenspace for λ = −3 is only onedimensional. We
conclude that A is nondiagonalizable, which suggests that we seek generalized
eigenvectors by solving (A −λI)
2
w = 0. However, because (A + 3I)
2
= 0, the
matrix
N = A +3I =
_
6 −18
2 −6
_
is nilpotent of order 2. This observation provides a useful decomposition for the
matrix A, namely A = −3I + N. Since the matrix −3I clearly commutes with N,
we are allowed to use the binomial theorem to calculate powers of A:
A
n
= (−3I + N)
n
=
n
∑
k=0
_
n
k
_
(−3I)
n−k
N
k
.
i ntroducti on to di fference equati ons 189
Since N is nilpotent of order 2, only the ﬁrst two terms in this sum survive:
A
n
=
_
n
0
_
(−3I)
n
N
0
+
_
n
1
_
(−3I)
n−1
N.
Here, the binomial coefﬁcients are (
n
0
) = 1 and (
n
1
) = n, and the matrix N
0
is
simply the identity matrix. Therefore,
A
n
=
_
(−3)
n
+6n(−3)
n−1
−18n(−3)
n−1
2n(−3)
n−1
(−3)
n
−6n(−3)
n−1
_
,
and the solution to the initial value problem is
_
x
n
y
n
_
= A
n
_
x
0
y
0
_
=
_
(−3)
n
+6n(−3)
n−1
−18n(−3)
n−1
2n(−3)
n−1
(−3)
n
−6n(−3)
n−1
_ _
1
1
_
=
_
(−3)
n
−12n(−3)
n−1
(−3)
n
−4n(−3)
n−1
_
.
If the coefﬁcient matrix associated with a constantcoefﬁcient system of dif
ference equations has complex conjugate eigenvalues, writing the closedform
solution can be messy. For example, consider the initial value problem
x
n+1
= −y
n
x
0
= 1
y
n+1
= x
n
y
0
= 1.
The coefﬁcient matrix
A =
_
0 −1
1 0
_
is in real canonical form. If this had been a system of odes, we would expect
periodic solutions involving sin t and cos t. In some sense, the behavior of this
discrete system is similar, as we can see by computing the ﬁrst few iterates:
_
x
1
y
1
_
=
_
−1
1
_ _
x
2
y
2
_
=
_
−1
−1
_
_
x
3
y
3
_
=
_
1
−1
_ _
x
4
y
4
_
=
_
1
1
_
.
190 li near, constant coeffi ci ent di fference equati ons
Evidently, this pattern will repeat, and the iterates will cycle through the
four different vectors shown here. Writing a closed formula for the solution is
straightforward, but doing so is a bit awkward because there are four cases to
consider.
In general, if the coefﬁcient matrix A has a pair of complex conjugate eigenval
ues α ± βi, it is useful to transform A into real canonical form:
A = P
_
α −β
β α
_
P
−1
,
where P is a suitablychosen invertible matrix. Letting M denote the real canoni
cal form for A, powers of A can be computed using A
n
= PM
n
P
−1
. To calculate
powers of M, it is helpful to split M as
M =
_
α −β
β α
_
= αI + B,
where
B =
_
0 −β
β 0
_
.
Certainly αI commutes with B, which allows us to use the binomial theorem
when computing M
n
= (αI + B)
n
. Moreover, since B
2
= −β
2
I, we may calculate
that B
3
= −β
2
B and B
4
= β
4
I. Since B
4
is a constant multiple of the identity
matrix, we might expect the same sorts of cyclic oscillations that we saw in the
example above.
Finally, we remark that solving inhomogeneous constantcoefﬁcient difference
equations is straightforward, but we shall not discuss the techniques here. Indeed,
for the inhomogeneous equation
x
n+1
= a
1
x
n
+ a
2
x
n−1
+ · · · + a
k
x
n−k+1
+ g(n),
it is possible to state an analogue of the variation of parameters formula (2.18)
for odes. Due to the discrete nature of difference equations, the solution contains
a summation involving g(n), as opposed to the integral in (2.18).
i ntroducti on to di fference equati ons 191
7.3. FirstOrder Nonlinear Equations and Stability
It is almost never possible to present a closed formula for the solution of a
nonlinear difference equation. As with nonlinear odes, we typically settle for
a qualitative understanding of how solutions behave. Our development of
the qualitative analysis of nonlinear difference equations perfectly parallels the
methodology we introduced for nonlinear odes. We will begin by analyzing
constant solutions, which are analogous to equilibria for odes. Afterwards,
we will study more exotic dynamical behavior, including periodic solutions,
bifurcations, and chaos. We restrict our initial discussion to ﬁrstorder nonlinear
difference equations, later generalizing our results to higherorder systems. The
material in this section is based heavily on Chapter 10 of Strogatz [11].
Example 7.3.1. The behavior of the iterates of the nonlinear equation x
n+1
= x
2
n
depends greatly upon our choice of initial condition x
0
. For example, if x
0
= 2,
then x
n
→ ∞ as n → ∞. On the other hand, if x
0
= 1/2, then the sequence of
iterates x
n
converges rapidly to 0. Notice also that if x
0
= 1, then x
n
= 1 for
all n ≥ 0. This constant solution of the difference equation is analogous to an
equilibrium for an ode, and such solutions have a special name.
Deﬁnition 7.3.2. A ﬁxed point of the ﬁrstorder difference equation x
n+1
= f (x
n
)
is any number x
∗
such that x
∗
= f (x
∗
).
Notice that, by deﬁnition, if we start out by using a ﬁxed point as our initial
condition, then we will remain stuck at that ﬁxed point for all future iterates.
Example 7.3.3. To ﬁnd all ﬁxed points x of the difference equation x
n+1
=
2x
n
−2x
2
n
, we should solve the equation x = 2x −2x
2
. By algebra, we have
2x(x −1/2) = 0, a quadratic equation with two roots: x = 0 and x = 1/2. These
are the two ﬁxed points of this nonlinear difference equation.
As with equilibria of odes, ﬁxed points of difference equations can be stable
or unstable. Roughly speaking, a ﬁxed point x
∗
is locally asymptotically stable if
whenever we start from an initial condition x
0
that is appropriately “close” to x
∗
,
the sequence {x
n
} of iterates converges to x
∗
as n →∞. Fixed points can also be
repellers—i.e., the gap between x
∗
and x
n
may grow as n increases, no matter
how close the initial condition x
0
is to x
∗
.
192 fi rst order nonli near equati ons and stabi li ty
Example 7.3.4. Fixed points of the mapping x
n+1
= x
2
n
satisfy the equation
x = x
2
. This quadratic equation has two solutions, x = 0 and x = 1. The ﬁxed
point x = 0 is locally asymptotically stable, because if we start from any initial
condition x
0
that is “close” to 0, then the sequence of iterates will converge to 0.
Speciﬁcally, if x
0
∈ (−1, 1), then x
n
→ 0 as n → ∞. In contrast, the ﬁxed point
x = 1 is unstable, because if we start from any initial condition other than x
0
= 1,
the iterates will be repelled from 1.
We now devise a test to determine whether a ﬁxed point of a difference
equation is locally stable or not. As with differential equations, the stability test
involves the use of Jacobian matrices. However, the conditions that eigenvalues
must satisfy will be different—stability will depend upon more than just the real
part of the eigenvalues.
First, consider the ﬁrstorder difference equation x
n+1
= f (x
n
), and assume
that the function f : R →R is continuously differentiable. Suppose that x
∗
is an
isolated ﬁxed point of our equation. To determine whether x
∗
is an attractor or
repeller, we need to investigate how the iterates of the mapping would behave
if we start from an initial condition that is “near” x
∗
. Suppose that our initial
condition is x
0
= x
∗
+
0
, where 
0
 is a very small number. We will estimate
the gap
1
between the value of x
1
(the ﬁrst iterate) and the ﬁxed point x
∗
in
order to see whether x
1
is closer to the ﬁxed point than x
0
was. More exactly,
suppose x
1
= x
∗
+
1
. We also know that x
1
= f (x
0
) and, since x
0
= x
∗
+
0
, we
may use the tangent line approximation at x
∗
to estimate
x
1
= f (x
∗
+
0
) ≈ f (x
∗
) +
0
f
/
(x
∗
).
Recalling that x
1
= x
∗
+
1
, we equate our two expressions for x
1
to obtain
x
∗
+
1
≈ f (x
∗
) +
0
f
/
(x
∗
).
The fact that x
∗
is a ﬁxed point implies that f (x
∗
) = x
∗
, and therefore
x
∗
+
1
≈ x
∗
+
0
f
/
(x
∗
).
Subtracting x
∗
from both sides and taking absolute values, we ﬁnd that
¸
¸
¸
¸
1
0
¸
¸
¸
¸
≈  f
/
(x
∗
).
i ntroducti on to di fference equati ons 193
Interpreting this approximation in words will give rise to our ﬁrst stability
criterion. Recall that
0
and
1
measure the gaps x
0
−x
∗
and x
1
−x
∗
, respectively.
Thus, the left hand side of the above approximation measures the ratio of these
gaps. If the ﬁxed point is an attractor, we would need this ratio to be smaller
than 1 in magnitude, implying that the gaps between iterates and the ﬁxed point
x
∗
will shrink as we generate more iterates. Conversely, if the ratio exceeds 1
in magnitude, then the gaps between the iterates x
n
and the ﬁxed point x
∗
will
grow as n increases. Thus, we have provided a heuristic proof of
Theorem 7.3.5. Suppose x
∗
is an isolated ﬁxed point of the ﬁrstorder difference
equation x
n+1
= f (x
n
), where f is continuously differentiable. Then x
∗
is locally
asymptotically stable (attracting) if  f
/
(x
∗
) < 1 and is unstable (repelling) if
 f
/
(x
∗
) > 1. If  f
/
(x
∗
) = 1, this test is inconclusive.
Warning: Although unstable ﬁxed points are locally repelling, we must exercise
caution when drawing conclusions about longterm behavior of iterates (particu
larly if f is not as smooth as required by the conditions of Theorem 7.3.5). If f is
merely piecewise continuous, it is possible for x
n+1
= f (x
n
) to have an unstable
ﬁxed point which is globally attracting (see exercises).
Example 7.3.6. Consider the difference equation x
n+1
= cos(x
n
). We claim that
this equation has exactly one ﬁxed point. Fixed points satisfy the transcendental
equation x = cos x, which is impossible to solve algebraically. Equivalently, ﬁxed
points are roots of the function g(x) = x −cos x. Notice that g(x) is a continuous
function and that g(0) = −1 whereas g(π/2) = π/2. Since g is continuous and
its values change from negative to positive between x = 0 and x = π/2, the
intermediate value theorem from calculus guarantees that g has at least one root
in the interval (0, π/2). Next, we must show that g has exactly one real root.
To see why, observe that g
/
(x) = 1 + sin x is nonnegative because sin x ≥ −1.
Thus, the function g(x) is nondecreasing (it is actually a onetoone function). It
follows that the equation g(x) = 0 can have at most one root. Letting x
∗
denote
this root, we conclude that x
∗
is the only ﬁxed point of this difference equation.
Again, it is impossible to ﬁnd the value of x
∗
algebraically. However, the above
remarks indicate that 0 < x
∗
< π/2, and this is actually enough information for
us to use Theorem 7.3.5 to test the local stability of x
∗
. Our difference equation
has the form x
n+1
= f (x
n
) where f (x) = cos x. According to the Theorem, we
should check the magnitude of f
/
(x
∗
). In this case, f
/
(x) = −sin x, from which
194 fi rst order nonli near equati ons and stabi li ty
we calculate
 f
/
(x
∗
) =  −sin(x
∗
).
Since 0 < x
∗
< π/2, we have  −sin x
∗
 < 1. Therefore, Theorem 7.3.5 guarantees
that our ﬁxed point x
∗
is locally asymptotically stable. In this example, we never
needed to know the exact value of x
∗
in order to test its stability.
Remarkably, the ﬁxed point x
∗
of this difference equation is actually globally
asymptotically stable. That is, for any choice of initial condition x
0
, the sequence
of iterates of this mapping will converge to the ﬁxed point! You should test this
out by picking any number you like and then using a calculator or computer to
repeatedly take the cosine of the number you chose. Make sure your calculator is
measuring angles in radians, not degrees. You will ﬁnd that the value of the ﬁxed
point is x
∗
= 0.739085..., which is the only solution of the equation x = cos x.
Example 7.3.7. Recall that the difference equation x
n+1
= x
2
n
has two ﬁxed
points, 0 and 1. In this case, f (x) = x
2
, so f
/
(x) = 2x. Since f
/
(0) = 0 < 1, we
see that 0 is a locally asymptotically stable ﬁxed point, and since f
/
(1) = 2 > 1,
we conclude that 1 is an unstable ﬁxed point.
Example 7.3.8. By algebra, you can check that the only ﬁxed points of x
n+1
=
3x
n
(1 − x
n
) are 0 and 2/3. Here, f (x) = 3x −3x
2
, so f
/
(x) = 3 −6x. Since
 f
/
(0) = 3 > 1, we see that 0 is an unstable ﬁxed point. On the other hand, since
 f
/
(2/3) = 1, we cannot use Theorem 7.3.5 to draw any conclusions regarding
the stability of that ﬁxed point.
Deﬁnition 7.3.9. A ﬁxed point x
∗
of the equation x
n+1
= f (x
n
) is called hyper
bolic if  f
/
(x
∗
) ,= 1. Otherwise, the ﬁxed point is called nonhyperbolic.
Our local stability Theorem 7.3.5 can only be used to classify stability of
hyperbolic ﬁxed points. In order to determine whether a nonhyperbolic ﬁxed
point is stable, we need a ﬁner approach. After all, the derivation of Theorem 7.3.5
was based upon linear approximation of the function f in the vicinity of a ﬁxed
point x
∗
. If f has a continuous third derivative, then we can obtain the following
theorems regarding stability of nonhyperbolic equilibria:
Theorem 7.3.10. Suppose that x
∗
is an isolated nonhyperbolic equilibrium point
of x
n+1
= f (x
n
) and, more speciﬁcally, that f
/
(x
∗
) = 1. Then x
∗
is unstable if
f
//
(x
∗
) ,= 0. If f
//
(x
∗
) = 0 and f
///
(x
∗
) > 0 then, again, x
∗
is unstable. Finally, if
f
//
(x
∗
) = 0 and f
///
(x
∗
) < 0, then x
∗
is locally asymptotically stable.
i ntroducti on to di fference equati ons 195
In order to state the corresponding theorem for the case f
/
(x
∗
) = −1, it is
helpful to introduce the notion of the Schwarzian derivative.
Deﬁnition 7.3.11. The Schwarzian derivative of a function f is deﬁned as
Sf (x) =
f
///
(x)
f
/
(x)
−
3
2
_
f
//
(x)
f
/
(x)
_
2
.
Theorem 7.3.12. Suppose that x
∗
is an isolated nonhyperbolic equilibrium point
of x
n+1
= f (x
n
) and that f
/
(x
∗
) = −1. Then x
∗
is unstable if Sf (x
∗
) > 0 and is
locally asymptotically stable if Sf (x
∗
) < 0.
Example 7.3.13. In our previous example, we found that x
∗
= 2/3 is a non
hyperbolic ﬁxed point of the difference equation x
n+1
= 3x
n
(1 − x
n
). Since
f (x) = 3x −3x
2
, we compute that the ﬁrst three derivatives of f are
f
/
(x) = 3 −6x, f
//
(x) = −6 and f
///
(x) = 0.
Since f
/
(x
∗
) = −1, we may use Theorem 7.3.12. The expression for the
Schwarzian derivative reduces to Sf (x
∗
) = −f
///
(x
∗
) −
3
2
f
//
(x
∗
)
2
, and we ﬁnd
that Sf (x
∗
) = −54 < 0. Theorem 7.3.12 tells us that the nonhyperbolic ﬁxed
point x
∗
= 2/3 is locally asymptotically stable.
In the preceding example, we were still able to classify the stability of the ﬁxed
point even though Theorem 7.3.5 was inconclusive. Usually, Theorems 7.3.5,
7.3.10 and 7.3.12 are enough to classify stability of ﬁxed points, although there
are cases in which all three theorems are inconclusive.
7.4. Systems of Nonlinear Equations and Stability
Fixed points for higherorder difference equations can be analyzed via techniques
that are very similar to the ones we developed when considering equilibria of
nonlinear systems of odes. To motivate our deﬁnition for ﬁxed points of higher
order equations, let us study a speciﬁc secondorder equation
x
n+1
= [3 + x
n
−x
n−1
]x
n
(1 −x
n
).
196 systems of nonli near equati ons and stabi li ty
As usual, we write this as a system of two ﬁrstorder equations by introducing a
new variable y
n
= x
n−1
. The resulting system is
x
n+1
= [3 + x
n
−y
n
]x
n
(1 −x
n
) and y
n+1
= x
n
.
A ﬁxed point of such a system should correspond to a constant solution of the
difference equation. In this example, a ﬁxed point would be any constant vector
(x
∗
, y
∗
) such that whenever (x
n
, y
n
) = (x
∗
, y
∗
), we have (x
n+1
, y
n+1
) = (x
∗
, y
∗
)
as well. More generally,
Deﬁnition 7.4.1. Suppose f : R
m
→R
m
and that x
n+1
= f (x
n
) is a system of m
ﬁrstorder difference equations. A ﬁxed point of the system is any vector x
∗
∈ R
m
such that x
∗
= f (x
∗
).
In the above example, we may solve for the ﬁxed points by setting x
n+1
=
x
n
= x
∗
and y
n+1
= y
n
= y
∗
. The equation y
n+1
= x
n
tells us that x
∗
= y
∗
, and
substituting this into the equation for x
n+1
yields x
∗
= 3x
∗
(1 − x
∗
). Solving
this quadratic equation yields x
∗
= 0 and x
∗
= 2/3 as the two solutions.
Consequently, the system has two ﬁxed points, (x
∗
, y
∗
) = (0, 0) and (x
∗
, y
∗
) =
(2/3, 2/3).
Example 7.4.2. Consider the nonlinear system
x
n+1
= 2x
n
−x
n
y
n
and y
n+1
= x
n
y
n
.
Fixed points (x, y) must simultaneously satisfy
x = 2x −xy and y = xy.
By algebra, these two equations are x(1 −y) = 0 and y(x −1) = 0. If x = 0 in
the ﬁrst of these, then the second would force y = 0 as well. Similarly, if y = 1 in
the ﬁrst equation, then this would force x = 1 in the second equation. We have
obtained precisely two equilibria, (x, y) = (0, 0) and (x, y) = (1, 1).
Stability. If x
∗
is a ﬁxed point of a system x
n+1
= f (x
n
) of m ﬁrstorder differ
ence equations, we need a way of testing the stability of x
∗
. We mimic exactly
what we did when linearizing systems of odes at an equilibrium point. Recall
that the linear approximation of a function f : R
m
→R
m
at a point x
∗
∈ R
n
is
i ntroducti on to di fference equati ons 197
given by
f (x) ≈ f (x
∗
) + J f (x
∗
)(x −x
∗
).
Using this approximation in the above difference equation,
x
n+1
≈ f (x
∗
) + J f (x
∗
)(x
n
−x
∗
) = x
∗
+ J f (x
∗
)(x
n
−x
∗
),
where we have used the fact that f (x
∗
) = x
∗
since x
∗
is a ﬁxed point. To measure
the gap between iterates of the difference equation and the ﬁxed point, deﬁne
y
n
= x
n
−x
∗
. Then in the vicinity of x
∗
, the vectors y
n
approximately satisfy
y
n+1
= J f (x
∗
)y
n
,
a linear, constantcoefﬁcient system. The exact solution of this linearized system
is given by
y
n
= [J f (x
∗
)]
n
y
0
,
where the vector y
0
is a measure of our initial gap x
0
−x
∗
. In order to ensure
that x
∗
is locally asymptotically stable, we need a criterion which guarantees
that the gap y
n
= x
n
−x
∗
will approach 0 as we let n → ∞. By inspecting the
solution of the linearized system, we must insist that the entries of the powers
of the Jacobian matrix J f (x
∗
) converge to 0 as n →∞. If this Jacobian matrix is
diagonalizable, we could write J f (x
∗
) = PDP
−1
where D is a diagonal matrix
containing the eigenvalues, and P is a constant, invertible matrix. The fact that
powers of D contain powers of the eigenvalues suggests a stability criterion: each
eigenvalue should have “size” smaller than 1 so that their powers will converge
to 0 as n →∞. For real numbers, the absolute value function gives us a notion
of “size”. We now generalize the concept of absolute value to include complex
numbers.
Deﬁnition 7.4.3. If z = x + iy is a complex number, then the modulus of z is
deﬁned as z =
_
x
2
+ y
2
.
Notice that if z is real (i.e., y = 0), then z =
√
x
2
= x and the modulus of z
is given by the usual absolute value function.
Now that we have a notion of “size” for complex numbers, all of our above
remarks constitute a heuristic proof of the following stability theorem.
198 systems of nonli near equati ons and stabi li ty
Theorem 7.4.4. Let f : R
m
→ R
m
and suppose that all entries in the Jacobian
matrix J f (x) are continuous. An isolated ﬁxed point x
∗
of a system x
n+1
= f (x
n
)
of m ﬁrstorder difference equations is
locally asymptotically stable if ALL eigenvalues of J f (x
∗
) have modulus less
than 1;
unstable if ANY eigenvalue of J f (x
∗
) has modulus greater than 1.
Remark. This stability criterion is different (although similar in spirit) from the
one we developed when analyzing equilibria of odes. An equilibrium x
∗
of
a system of odes is stable if all eigenvalues of J f (x
∗
) have negative real part.
By contrast, a ﬁxed point x
∗
for a system of difference equations is stable if all
eigenvalues of J f (x
∗
) have modulus less than 1.
Example 7.4.5. Find all ﬁxed points of the system
x
n+1
= 3x
n
−x
n
y
n
y
n+1
= −2y
n
+ x
n
y
n
and determine whether they are locally stable or unstable.
Solution: Fixed points (x, y) of this system satisfy
x = 3x −xy and y = −2y + xy.
Equivalently, x(2 − y) = 0 and y(x − 3) = 0. If x = 0 in the ﬁrst of these
equations, then we are forced to set y = 0 in order to satisfy the second equation.
Likewise, if y = 2 in the ﬁrst equation, then we must have x = 3 in the second
equation. Thus, there are two ﬁxed points, (0, 0) and (3, 2). For this system, the
function f : R
2
→R
2
is deﬁned by
f (x, y) =
_
3x −xy
−2y + xy
_
,
and its Jacobian matrix is
J f (x, y) =
_
3 −y −x
y x −2
_
.
i ntroducti on to di fference equati ons 199
At the ﬁxed point (0, 0), the Jacobian matrix J f (0, 0) is a diagonal matrix, namely
diag{3, −2}, and the eigenvalues are the diagonal entries. The eigenvalues
are real, so their moduli are simply their absolute values. Since there is an
eigenvalue with modulus larger than 1 (in fact both have modulus larger than 1),
Theorem 7.4.4 tells us that (0, 0) is an unstable equilibrium.
At the other ﬁxed point (3, 2), the Jacobian matrix is
J f (x, y) =
_
1 −3
2 1
_
,
and the associated characteristic equation is λ
2
−2λ +7 = 0. From the quadratic
formula, the eigenvalues are λ
±
= 1 ±
√
6i. These two eigenvalues have
modulus
λ
±
 =
_
1
2
+ (
√
6)
2
=
√
7 > 1.
Again, we conclude that the ﬁxed point is unstable because there are eigenvalues
with modulus exceeding 1.
For 2 2 matrices, there is actually a quick way to determine whether all
eigenvalues have modulus less than 1 without actually computing the eigenval
ues. The following Lemma can be used to test the stability of ﬁxed points for
systems of two ﬁrstorder difference equations.
Lemma 7.4.6. Suppose that A is a 2 2 matrix. Then the eigenvalues of A have
modulus less than 1 if and only if
trA +det A > −1, trA −det A < 1, and det A < 1.
The proof of Lemma 7.4.6 is relatively straightforward after you recall how to
express the eigenvalues of A in terms of the trace and determinant. This Lemma
is a special case of the Jury Stability Test, which provides an iterative procedure
by which one may determine whether all eigenvalues of an n n matrix have
modulus less than 1.
Example 7.4.7. The matrix
A =
_
1 −1
1
4
−1
_
200 peri od doubli ng bi furcati ons
has trace 0 and determinant −3/4. All three criteria of Lemma 7.4.6 are satisﬁed,
which means that the eigenvalues of A have modulus less than 1.
Lemma 7.4.6 tends to be especially useful for systems of difference equations
which contain unspeciﬁed parameters, because it can provide inequalities that
the parameters must satisfy so that equilibria will be stable. Such systems are
the subject of the following section.
7.5. PeriodDoubling Bifurcations
In a previous chapter, we discussed how the dynamical behavior of systems
of odes can exhibit sudden, dramatic changes as a parameter µ is varied. The
same can be said for systems of difference equations involving a parameter.
Undoubtedly the most famous example of this sort is the discrete logistic equation,
a ﬁrstorder difference equation involving a single parameter µ. This difference
equation is given by
x
n+1
= µx
n
(1 −x
n
), (7.1)
where µ is a real number. From now on, we will insist that 0 ≤ µ ≤ 4, because
this restriction on µ endows Equation (7.1) with a convenient property. Namely,
if the initial condition x
0
is in the interval [0, 1], then all subsequent iterates
will remain in the interval [0, 1]. To see why, simply inspect the graph of the
function f (x) = µx(1 −x). If µ > 0, then the graph of f (x) is a parabola with
roots at x = 0 and x = 1. The maximum value of f occurs at x = 1/2, and the
value of f (x) at its maximum is µ/4. Thus, if 0 ≤ µ ≤ 4 and 0 ≤ x ≤ 1, then
0 ≤ f (x) ≤ 1.
Fixed points. The discrete logistic equation has two ﬁxed points. To ﬁnd them,
we set x
n+1
= x
n
= x in Equation (7.1), obtaining x = µx(1 −x). One solution
is x = 0, which is a ﬁxed point independent of the value of µ. The other ﬁxed
point is x = 1 −1/µ which does, of course, change as the parameter µ changes.
Theorem 7.3.5 can be used to determine whether these two ﬁxed points are
locally stable. The righthand side of the logistic equation is deﬁned by the
function f (x) = µx(1 − x) = µx −µx
2
. Taking the derivative f
/
(x) = µ −2µx,
we compute that  f
/
(0) = µ. According to the theorem, we need µ < 1 in
order to ensure the stability of the ﬁxed point x = 0. Since we have already
assumed that 0 ≤ µ ≤ 4, we conclude that x = 0 is locally asymptotically stable if
i ntroducti on to di fference equati ons 201
0 ≤ µ < 1 and is unstable if 1 < µ ≤ 4. For the other ﬁxed point x = 1 −1/µ,
we calculate that f
/
(1 −1/µ) = 2 −µ. Again, Theorem 7.3.5 will guarantee
that the ﬁxed point is stable provided that 2 −µ < 1. Equivalently, we have
shown that x = 1 −1/µ is locally asymptotically stable if 1 < µ < 3, and is unstable
if either µ < 1 or µ > 3.
For µ > 3, something very interesting happens. Both x = 0 and x = 1 −1/µ
are unstable ﬁxed points, and they act as repellers. However, we have already
shown that whenever we start from an initial condition x
0
∈ [0, 1], then all
subsequent iterates x
n
are trapped inside the interval [0, 1]. How is the logistic
equation (7.1) able to resolve the fact that the iterates x
n
are trapped inside an
interval that contains two repelling ﬁxed points? Apparently, if we start from
any initial condition other than one of the two ﬁxed points, then the system can
never
1
“settle down” to equilibrium.
Suppose we choose µ = 3.1 and pick any initial condition (other than one of
the two ﬁxed points) and use a computer to generate a list of the ﬁrst few iterates.
It turns out that the iterates converge to an alternating pattern—i.e., the sequence
of iterates with even index x
2n
converges to one number, whereas the sequence
of oddindexed iterates x
2n+1
converges to a different number. The bifurcation
that occurs when µ = 3 gives rise to alternation, and the solutions that we obtain
are analogous to periodic solutions for odes. For onedimensional difference
equations, when a ﬁxed point x
∗
loses stability because f
/
(x
∗
) decreases past
−1, we say that a perioddoubling bifurcation has occurred. The discrete logistic
equation (7.1) experiences a perioddoubling bifurcation when µ = 3. For the
speciﬁc choice µ = 3.1, the iterates eventually alternate between approximately
0.558014 and 0.764567, assuming that our initial condition x
0
does not coincide
with one of the two unstable ﬁxed points. This raises a natural question: “What
are these numbers?”
The answer to this question is actually very sensible after we interpret what
it would mean for every second iterate of the difference equation to be the
same. Mathematically, it means that x
n+2
= x
n
for each n. Moreover, since
x
n+2
= f (x
n+1
) = f ( f (x
n
)), we are led to explore the second iterate mapping
x
n+2
= f ( f (x
n
)) = µ[µx
n
(1 −x
n
)][1 −µx
n
(1 −x
n
)]. (7.2)
1
Read the warning that appears after Theorem 7.3.5.
202 peri od doubli ng bi furcati ons
Apparently, the two mystery numbers listed above are ﬁxed points of the
second iterate mapping, but not of the original logistic equation (7.1). From
Equation (7.2), ﬁxed points x of the second iterate mapping must satisfy the
algebraic equation
x = µ[µx(1 −x)][1 −µx(1 −x)],
a fourthdegree equation for x. Solving this equation appears daunting until we
observe that any ﬁxed point of the original equation (7.1) is deﬁnitely a ﬁxed
point of the second iterate mapping (7.2) as well. Therefore, we already know
that x = 0 and x = 1 −1/µ are solutions of the above fourthdegree equation.
After tedious factorization, we ﬁnd that the other two roots of this equation are
x
±
=
(µ +1) ±
_
(µ −3)(µ +1)
2µ
, (7.3)
which are real numbers provided that µ > 3. Both of these are ﬁxed points
of the second iterate mapping but not of the logistic equation itself. We have
noted that for µ slightly larger than 3, the iterates of the logistic equation settle
into an alternating pattern, sometimes known as a period2 solution or a 2cycle.
Moreover, formula (7.3) tells us the two numbers that the iterates will alternate
between. Interestingly, the 2cycle is locally asymptotically stable: if we start from
any initial condition that is “near” either x
+
or x
−
, the iterates of the logistic
equation will always converge to the 2cycle solution. Testing the stability of
a periodic solutions (cycles) of a difference equation is actually much more
straightforward than testing stability of periodic solutions for odes. We know
that if x
∗
is a ﬁxed point of a ﬁrstorder difference equation x
n+1
= f (x
n
), then
x
∗
is locally asymptotically stable if  f
/
(x
∗
) < 1. We can apply the same criterion
to the second iterate mapping (7.2): Consider the ﬁxed point x
−
of the equation
x
n+2
= f ( f (x
n
)). If we deﬁne g(x) = f ( f (x)), then the stability criterion would
require that g
/
(x
−
) < 1. By the chain rule, g
/
(x
−
) = f
/
( f (x
−
)) f
/
(x
−
). Finally,
since the iterates in the 2cycle alternate between x
−
and x
+
, we know that
f (x
−
) = x
+
. It follows that the 2cycle of the discrete logistic equation is locally
asymptotically stable if
 f
/
(x
+
) f
/
(x
−
) < 1.
After tedious algebra, this stability condition reveals that the 2cycle solution is
locally asymptotically stable if 3 < µ < 1 +
√
6. As µ increases beyond 1 +
√
6,
the two ﬁxed points x
±
of the second iterate mapping (see Equations (7.2)
i ntroducti on to di fference equati ons 203
and (7.3) above) lose their stability via another perioddoubling bifurcation. The
2cycle deﬁned by x
±
still exists for µ > 1 +
√
6, but it is unstable. The period
doubling bifurcation at µ = 1 +
√
6 creates a stable 4cycle. More exactly, as
soon as the ﬁxed points of the second iterate mapping lose their stability, the
fourth iterate mapping x
n+4
= f ( f ( f ( f (x
n
)))) gains four stable ﬁxed points. For
µ values slightly larger than 1 +
√
6, iterates of the discrete logistic mapping
will repeatedly cycle through these four values. Notice that since the function
f (x) = µx(1 −x) is quadratic, then composing it with itself four times results in
an 8thdegree polynomial equation. Solving for the ﬁxed points of the fourth
iterate mapping is possible, but is certainly not easy.
Further increasing µ, the cascade of perioddoubling bifurcations occurs with
increasing frequency. In fact, M.J. Feigenbaum was able to prove that the
lengths of intervals between successive perioddoubling bifurcations decrease to
0 geometrically in the following sense: If µ
n
is the value of µ at which a period2
n
cycle ﬁrst occurs, then
lim
n→∞
µ
n
−µ
n−1
µ
n+1
−µ
n
= 4.669201609102990 . . .
Since the intervals between successive bifurcations decrease approximately geo
metrically by this factor, it follows that there should be some critical value of µ
at which the period of the cycle becomes inﬁnite. Indeed this is the case, and it
occurs when µ is approximately 3.569946. At that critical µ, the behavior of the
iterates of the discrete logistic mapping becomes chaotic (see next section).
Bifurcation diagram. The cascade of perioddoubling bifurcations in the logistic
mapping is easiest to visualize via a bifurcation diagram. The idea is to plot the
longterm stable response of equation (7.1) versus the parameter µ. The bifurcation
diagram appears in Figure 7.1, and we now summarize how to interpret the
diagram. For 0 ≤ µ < 1, we know that 0 is a stable ﬁxed point of the logistic
equation, and there are no other stable ﬁxed points. For 1 < µ < 3, the points
in the bifurcation diagram follow the curve x = 1 −1/µ, since 1 −1/µ is a
stable ﬁxed point for that range of µ values. At µ = 3, the curve branches into
two curves which persist until µ = 1 +
√
6. Given any µ ∈ (3, 1 +
√
6), the
longterm behavior of the iterates is alternation between the two branches, which
correspond to the values of x
±
deﬁned in (7.3). For µ slightly larger than 1 +
√
6,
there are four “branches” in the bifurcation diagram. If we select µ from this
204 chaos
x
0 1 2 3
4
0
1
µ
.
Figure 7.1. Bifurcation diagram for the discrete logistic equation (7.1).
“period4 window”, the iterates of the discrete logistic equation will (eventually)
cycle repeatedly through the four different values indicated by the branches. For
most choices of µ > 3.569946, the branches in the bifurcation diagram become a
blur. Apparently, the discrete logistic equation exhibits erratic behavior in that
parameter regime (see next section). It is interesting to note that there is actually
a period3 cycle in the vicinity of µ = 3.83, which can be seen if you look closely
at the ﬁgure.
7.6. Chaos
For most values of µ > 3.57, the iterates of the discrete logistic equation (7.1)
form an aperiodic sequence. The fact that this simple equation can produce
such erratic behavior without relying upon artiﬁcial “randomness” is rather
remarkable. In fact, the discrete logistic equation is, in some sense, the simplest
equation which exhibits chaotic dynamics. Let us recall the deﬁnition of chaos,
which was stated in a previous chapter. A wellwritten, concise deﬁnition of
chaos is provided in the text of Strogatz [11]; we repeat the deﬁnition exactly as it
appears in his text:
Deﬁnition 7.6.1. “Chaos is aperiodic longterm behavior in a deterministic system
that exhibits sensitive dependence on initial conditions.”
i ntroducti on to di fference equati ons 205
There are three key phrases that appear in this deﬁnition. By aperiodic longterm
behavior, we typically mean that the system has solutions that remain bounded
but never converge to a ﬁxed point or periodic orbit. By deterministic, we mean
that the system is not allowed to incorporate any randomness in its parameters
and/or inputs. The system must be able to produce erratic, aperiodic behavior on
its own—artiﬁcially including randomness would be cheating. Finally, sensitive
dependence on initial conditions means that if we start from two different initial
conditions that are “nearby”, then the corresponding solutions trajectories will
separate exponentially fast (at least in the short term).
Testing whether a sequence of numbers is periodic can be quite challeng
ing, because recognizing repeating patterns is difﬁcult when the period of the
oscillations is large. Fortunately, there are standard techniques for testing for pe
riodicity within a sequence {x
n
} of real numbers. The discrete Fourier transform
and, more speciﬁcally, the fast Fourier transform (fft) are examples of such
techniques. The fft is an extremely useful way of determining the dominant
frequencies of oscillations that may be hidden within sequences of numbers. We
shall not cover the fft in this text; the interested reader is encouraged to refer to
any text on Fourier analysis, such as [1]. Instead, we focus on the problem of
quantifying sensitive dependence on initial conditions.
Lyapunov exponents. We now devise a test for sensitive dependence on
initial conditions for ﬁrstorder difference equations. Consider the equation
x
n+1
= f (x
n
), where f is continuously differentiable, and let x
0
denote an initial
condition. In order to measure the discrepancy between two nearby solution
trajectories, suppose we use the same difference equation y
n+1
= f (y
n
) with
a different initial condition y
0
= x
0
+
0
, where 
0
 is small and positive. If
0
= y
0
− x
0
represents the initial gap between our two solutions, we need
to estimate the gap
n
= y
n
− x
n
after n iterations. Note that x
n
= f
(n)
(x
0
),
where f
(n)
denotes the nth iterate mapping—i.e., f
(n)
represents the function
f composed with itself n times, not the nth derivative of f . Similarly, y
n
=
f
(n)
(y
0
) = f
(n)
(x
0
+
0
), which implies that
n
= f
(n)
(x
0
+
0
) − f
(n)
(x
0
).
Dividing by the initial gap
0
, we obtain the approximation
n
0
=
f
(n)
(x
0
+
0
) − f
(n)
(x
0
)
0
≈ [ f
(n)
]
/
(x
0
). (7.4)
206 chaos
(The last approximation would be an equality if we let
0
→ 0.) The derivative
of the function f
(n)
at x
0
is actually relatively straightforward to compute:
Lemma 7.6.2. Assume that x
n+1
= f (x
n
), where f is continuously differentiable
and let f
(n)
denote the composition of f with itself n times. Then
[ f
(n)
]
/
(x
0
) =
n−1
∏
k=0
f
/
(x
i
).
The product notation Π is similar to the summation notation Σ, except that
terms are multiplied rather than summed.
Proof. It is instructive to start with the case n = 2, in which case f
(2)
(x) =
f ( f (x)). Taking the derivative via the chain rule, we have
d
dx
f ( f (x)) = f
/
( f (x)) f
/
(x).
If we set x = x
0
, note that f (x
0
) = x
1
. Therefore, our expression becomes
[ f
(2)
]
/
(x
0
) = f
/
(x
1
) f
/
(x
0
).
The remainder of the proof is a straightforward induction on n.
With this Lemma in mind, Equation (7.4) can be rewritten as
n
0
≈
n−1
∏
k=0
f
/
(x
i
). (7.5)
Since we wish to determine whether solutions of our nearby initial value prob
lems separate exponentially fast, suppose that 
n
 ≈
0
e
λn
for some constant
λ. If the gap
n
= y
n
−x
n
obeys such a relationship, then taking absolute values
in Equation (7.5) yields
e
λn
≈
¸
¸
¸
¸
¸
n−1
∏
k=0
f
/
(x
i
)
¸
¸
¸
¸
¸
.
Taking logarithms, we exploit the fact that the logarithm of a product is the sum
of the individual logarithms:
λ ≈
1
n
ln
¸
¸
¸
¸
¸
n−1
∏
k=0
f
/
(x
i
)
¸
¸
¸
¸
¸
=
1
n
n−1
∑
k=0
ln  f
/
(x
i
).
i ntroducti on to di fference equati ons 207
This calculation motivates the following deﬁnition:
Deﬁnition 7.6.3. If the limit
λ = lim
n→∞
1
n
n−1
∑
k=0
ln  f
/
(x
i
)
exists, it is called the Lyapunov exponent for the solution with initial condition x
0
.
There are two key things to notice about the deﬁnition of the Lyapunov
exponent λ. First, our deﬁnition of λ was based upon an assumed relationship
between the separation
n
= y
n
− x
n
and the initial separation
0
—namely,

n
 = 
0
e
λn
. If we calculate that the Lyapunov exponent λ is positive for some
initial condition x
0
, this suggests that if we switch to a different initial condition
near x
0
, the new solution trajectory will separate exponentially fast from the
original one. In other words,
A positive Lyapunov exponent λ is an indicator of chaos.
Second, notice that λ depends upon the choice of the initial condition x
0
.
Estimating a Lyapunov exponent by hand is typically not possible, so one
typically uses a computer to estimate the value of the inﬁnite sum. However,
there are some special cases worth mentioning. Suppose that x
∗
is a stable,
hyperbolic ﬁxed point of the difference equation x
n+1
= f (x
n
), in which case
we know that  f
/
(x
∗
) < 1. If we start with the initial condition x
0
= x
∗
, then
x
n
= x
∗
for all n ≥ 0 as well. Finally, since  f
/
(x
∗
) < 1, then ln  f
/
(x
∗
) < 0 and
therefore
λ = lim
n→∞
1
n
n−1
∑
k=0
ln  f
/
(x
∗
) = lim
n→∞
1
n
_
n ln  f
/
(x
∗
)
¸
= ln  f
/
(x
∗
) < 0.
It is not surprising that we get a negative Lyapunov exponent if we start at a
stable ﬁxed point as our initial condition—a constant solution that attracts nearby
solution trajectories is certainly not chaotic.
For the discrete logistic equation (7.1), it is interesting to see how the Lyapunov
exponent changes as we increase the parameter µ, starting from a random initial
condition x
0
∈ (0, 1). We would expect the Lyapunov exponent to satisfy λ ≤ 0
for µ < 3.5699, prior to the onset of chaos. Indeed, this is the case as illustrated
in Figure 7.2. Notice that the value of λ is actually equal to 0 at each µ value
at which a perioddoubling bifurcation occurs, such as µ = 3 and µ = 1 +
√
6.
208 how to control chaos
µ = 3.0 1+ 6 µ =
2.8 3.1 3.4 3.7
4.0
−1.5
0
1.0
λ
µ
.
Figure 7.2. Lyapunov exponents for the discrete logistic equation (7.1).
The ﬁrst instance of a positive Lyapunov exponent occurs at µ = 3.5699..., an
indication of chaos. In the vicinity of µ = 3.83, the value of λ actually becomes
negative again. For µ values in that region, the logistic equation exhibits stable,
period3 behavior.
Deterministic chaos is a truly fascinating phenomenon, and readers interested
in a continued development of chaos theory are encouraged to read Chapter 10
of Strogatz [11]. Despite the allure of chaos, it can be quite devastating when
a system begins to behave chaotically. For example, the spatial and temporal
patterns of the electrical waves in ﬁbrillating heart ventricles (an oftenfatal
arrhythmia) are chaotic. We conclude this Chapter with a discussion of a
feedback control algorithm for preventing the occurrence of chaos.
7.7. How to Control Chaos
The subject of chaos control was popularized in the 1990s by the work of Ott,
Grebogi and Yorke. In this section, we introduce a rather intuitive chaos control
algorithm known as timedelay autosynchronization (tdas), which was introduced
via the theory and experiments of Socolar et al. [9].
The assumption underlying tdas control is that by making small adjustments
to a parameter during each iteration, we can convert an unstable ﬁxed point into
i ntroducti on to di fference equati ons 209
an attractor. Now suppose that we have a system that behaves chaotically, such
as the discrete logistic equation for µ appropriately chosen. In the midst of the
chaotic iterates, there may be unstable ﬁxed points. In a physical system, we
would never actually see these ﬁxed points—for example, suppose that you try
to balance a broomstick on the palm of your hand, keeping your hand perfectly
still. If you could position the broomstick perfectly, it would stay in the vertical
position (analogous to an unstable equilibrium). Applying tdas control would
be analogous to making tiny movements of your hand in such a way that the
broom remains in the vertical position.
For a more speciﬁc description of tdas, consider the discrete logistic equation
x
n+1
= µx
n
(1 − x
n
) where µ is chosen from the interval 3 < µ < 1 +
√
6. For
that range of µ values, we know that the longterm behavior of the system
will be a period2 cycle, with the iterates alternating between two numbers
x
±
. Somewhere between these two iterates lies the unstable ﬁxed point x
∗
=
1 −1/µ. The tdas algorithm makes tiny adjustments to the system parameter µ
during each iteration, in such a way that the period2 cycle is terminated and x
∗
becomes an attractor. Speciﬁcally, suppose that we modify µ, replacing it with
µ + γ(x
n
−x
n−1
), where γ is a positive number. The modiﬁed logistic equation
is
x
n+1
= [µ + γ(x
n
−x
n−1
)]x
n
(1 −x
n
). (7.6)
Notice that we have adjusted the value of µ by an amount proportional to
the gap between the previous two iterates. If x
n
> x
∗
> x
n−1
, we know that
for the original logistic equation, x
n+1
would be smaller than x
∗
because the
iterates will alternate largesmalllargesmall. The modiﬁed logistic equation (7.6)
effectively adjusts µ by γ(x
n
− x
n−1
), which would be positive if x
n
> x
n−1
.
Hence, the modiﬁed logistic equation would compute a larger value for x
n+1
than the original logistic equation would have. Similarly, if x
n
< x
∗
< x
n−1
, then
the original logistic equation (7.1) would generate a value of x
n+1
which is larger
than x
∗
. The modiﬁed logistic equation (7.6) would adjust µ by γ(x
n
− x
n−1
),
which is negative in this case. Thus, the value of x
n+1
generated by (7.6) should
be smaller (and presumably closer to the ﬁxed point x
∗
) than the value of x
n+1
generated by (7.1).
This strategy is analogous to “robbing from the rich and giving to the poor”.
If the iterates are alternating largesmall and x
n+1
is expected to be large, then
210 how to control chaos
we will rob from it by effectively reducing µ. Likewise, if x
n+1
is expected to be
small, then we will give to it by effectively raising µ.
Deﬁnition 7.7.1. The above technique of modifying a system parameter µ by an
amount proportional to the previous two iterates is called timedelay autosynchro
nization (tdas).
Stability analysis. Above we suggested that the tdas algorithm can sometimes
successfully convert an unstable ﬁxed point into a stable one. Let us investigate
how this takes place for the discrete logistic equation. With tdas control, we
found that the logistic equation becomes
x
n+1
= [µ + γ(x
n
−x
n−1
)]x
n
(1 −x
n
).
Observe that the introduction of the term γ(x
n
−x
n−1
) has absolutely no effect
on the values of the ﬁxed points. Indeed, if x
∗
is one of the two ﬁxed points of
the original logistic equation, then x
∗
will still be a ﬁxed point of the modiﬁed
equation because γ(x
∗
− x
∗
) = 0. Equation (7.6) is a secondorder difference
equation, so we write it as a system in the usual way by letting y
n+1
= x
n
and
y
n
= x
n−1
:
x
n+1
= f
1
(x, y) = [µ + γ(x
n
−y
n
)]x
n
(1 −x
n
)
y
n+1
= f
2
(x, y) = x
n
.
(7.7)
Since x
∗
= 0 and x
∗
= 1 −1/µ were the ﬁxed points of the original logistic
equation, it follows that (x, y) = (0, 0) and (x, y) = (1 −1/µ, 1 −1/µ) are the
ﬁxed points of (7.7). We are particularly interested in the stability of the latter
ﬁxed point, which corresponds to the more interesting ﬁxed point of the original
logistic mapping (7.1).
The Jacobian matrix associated with the right hand side of (7.7) is
J f (x, y) =
_
µ −2µx +2γx −γy −3γx
2
+2γxy −γx + γx
2
1 0
_
.
Evaluating the Jacobian at the ﬁxed point (1 −1/µ, 1 −1/µ) yields
J f (1 −1/µ, 1 −1/µ) =
_
_
2 −µ + γ
_
1
µ
−
1
µ
2
_
γ
_
1
µ
2
−
1
µ
_
1 0
_
_
.
i ntroducti on to di fference equati ons 211
We are now in a position to use the Jury stability test (Lemma (7.4.6)) to determine
whether there exists a range of γ values which would make our ﬁxed point stable.
Letting A denote our Jacobian matrix, the trace and determinant are given by
tr(A) = 2 −µ + γ
_
1
µ
−
1
µ
2
_
and det(A) = γ
_
1
µ
−
1
µ
2
_
.
According to Lemma (7.4.6), there are three conditions that must be satisﬁed to
ensure stability. First, we need tr(A) −det(A) < 1, which imposes the constraint
µ > 1. This actually is not restrictive at all, because there is no need to attempt
tdas for µ ≤ 1. Next, we need det(A) < 1. By algebra, this leads to an inequality
that γ must satisfy, namely
γ <
µ
2
µ −1
.
The right hand side of this inequality is forced to be positive because of the other
constraint that µ > 1. Finally, we require that tr(A) + det(A) > −1, which by
algebra imposes the constraint
γ >
µ
2
(µ −3)
2(µ −1)
.
This lower bound on γ only becomes interesting when µ > 3, because our ﬁxed
point is certainly stable for 1 < µ < 3.
In order to determine whether the tdas scheme may successfully stabilize
(1 −1/µ, 1 −1/µ) when µ > 3, we must determine whether it is possible to
simultaneously satisfy both of the above inequalities for γ; i.e.,
µ
2
(µ −3)
2(µ −1)
< γ <
µ
2
µ −1
.
Figure 7.3 shows that it is, indeed, possible to satisfy these inequalities even in
the regime of µ values where chaos occurs. The ﬁgure shows a plot of the two
curves deﬁned by the above inequality on γ. Notice that the lower boundary is 0
when µ = 3, which makes sense because our ﬁxed point was already stable when
µ < 3. As µ increases, the lower boundary of the “control domain” increases,
indicated that larger γ would be necessary to control the response. In fact, as µ
approaches 4.0, the ﬁgure suggests that we must choose γ considerably larger
than 2 in order to successfully stabilize our ﬁxed point. The range of γ for which
212 how to control chaos
control is predicted to succeed becomes narrower as µ increases. However, even
in the regime where chaos exists (µ > 3.57), there are substantial ranges of γ for
which control may be successful. The ﬁgure indicates that if µ = 3.65, which
is well within the chaos zone, then using γ = 3.0 should be enough for tdas
to terminate the chaos. However, below we will point out a ﬂaw in this line of
reasoning.
Figure 7.4 illustrates the use of tdas control to terminate the period2 response
in the discrete logistic equation with µ = 3.2. The ﬁrst 20 iterates were computed
by iterating equation (7.1) (no control), resulting in the alternating pattern. The
next 20 iterates were computed from Equation (7.6), using γ = 1.5, to simulate
the effect of tdas control. After a very brief transient, the iterates converge to
the ﬁxed point 1 −1/µ = 1 −1/3.2. In the absence of control, this ﬁxed point
had been unstable.
Bad news: Although Figure 7.3 suggests that tdas should succeed for a wide
range of γ, unfortunately the ﬁgure gives a far too optimistic prediction of the
control domain. There are two notable reasons why this is the case. First, the
tdas method adjusts the parameter µ based only upon the two most recent iterates
of the underlying difference equation. In the chaos regime (µ > 3.5699), the
iterates of the equation behave so erratically that we should really incorporate
more “history” when deciding how µ should be adjusted. Second, Figure 7.3
was generated based upon a local stability analysis. Again, the erratic dynamical
behavior for µ > 3.5699 can land us in trouble if we blindly apply the tdas
technique. If γ is reasonably large, it is certainly possible for γ(x
n
−x
n−1
) to be
so large that the computed value of x
n+1
lands outside of the interval [0, 1]. This
causes the iterates to behave very badly for the discrete logistic mapping.
Good news: There is an improved version of tdas known as extended tdas, or
etdas, which is bettersuited for chaos control (see Socolar et al. [9]). Whereas
tdas modiﬁes µ by an amount proportional to two preceding iterates, etdas
incorporates many previous iterates.
More good news: In the example of the discrete logistic mapping, we have the
luxury of knowing the exact value of the unstable ﬁxed point. Amazingly, neither
tdas nor etdas control methods require us to actually know the location of the
unstable ﬁxed point(s) in advance. Simply put, these methods can ﬁnd unstable
i ntroducti on to di fference equati ons 213
3.25 3.5 3.75 4.0
µ
γ
4.5
6.0
3.0
1.5
0
control fails
control
fails
control succeeds
.
3.0
Figure 7.3. Predicted domain in which tdas control succeeds for the
discrete logistic equation (7.1).
x
n
1
0
20 30 40 10 0
n
control off control on
.
Figure 7.4. Illustration of tdas control in the discrete logistic mapping
using µ = 3.2 and γ = 1.5. The ﬁrst 20 iterates show the stable period2
response of the logistic equation (7.1). To simulate the onset of tdas
control, the next 20 iterates are computed using Equation (7.6), which
includes control.
ﬁxed points by reversing their stability, eliminating any periodic and/or chaotic
behavior in the process.
Exercises
1. Find a closed formula for the solution of x
n+1
= 2x
n
+3x
n−1
, x
0
= 1, x
1
= 0.
214 how to control chaos
2. Find a closed formula for the solution of x
n+1
= −4x
n
−4x
n−1
, x
0
= 1, x
1
= 2.
3. Find a closed formula for the solution of x
n+1
= −4x
n−1
, x
0
= 1, x
1
= 2.
Notice that this is a secondorder difference equation. The coefﬁcient of x
n
happens to be 0.
An alternate approach for solving constantcoefﬁcient equations: In your
ﬁrst course on odes, you probably studied linear, homogeneous constant
coefﬁcient problems such as y
//
+ 5y
/
− 6y = 0. If you seek exponential
solutions of the form y = e
λt
, you obtain a characteristic equation λ
2
+5λ −
6 = 0, which has distinct, real roots λ = −6 and λ = 1. This implies that
y = e
−6x
and y = e
x
are solutions of the ode and, since they are linearly
independent, the general solution is y = C
1
e
−6x
+ C
2
e
x
. On the other hand,
when the characteristic equation has a real root λ with multiplicity k, then we
expect solutions of the form e
λt
, te
λt
, . . . t
k−1
e
λt
.
In Problems 4 and 5, you will apply the same approach to solve linear,
homogeneous, constantcoefﬁcient difference equations.
4. Consider the initial value problem
x
n+1
= −2x
n
+8x
n−1
, x
0
= 5, x
1
= −2.
(a) Suppose we seek solutions of the form x
n
= λ
n
, excluding the unin
teresting case λ = 0. Show that λ must satisfy a characteristic equation
λ
2
+2λ −8 = 0.
(b) Note that the roots of this characteristic equation are real and distinct. By
drawing an analogy with the above ode example, write down the general
solution of this difference equation.
(c) Finally, use the initial conditions to ﬁnd the closed formula for x
n
. Ver
ify that your solution satisﬁes the difference equation and both initial
conditions.
5. Consider the initial value problem
x
n+1
= 6x
n
−9x
n−1
, x
0
= 2, x
1
= 3.
(a) Suppose we seek solutions of the form x
n
= λ
n
, excluding the unin
teresting case λ = 0. Show that λ must satisfy a characteristic equation
λ
2
−6λ +9 = 0.
i ntroducti on to di fference equati ons 215
(b) Note that λ = 3 is a repeated real root of the characteristic equation. By
analogy with odes, we not only expect solutions of the form x
n
= λ
n
, but
also solutions of the form x
n
= nλ
n
. Use this intuition to write down the
general solution of the difference equation.
(c) Finally, use the initial conditions to ﬁnd the closed formula for x
n
. Ver
ify that your solution satisﬁes the difference equation and both initial
conditions.
6. Find all ﬁxed points of x
n+1
= x
3
n
and classify their stability.
7. Consider the difference equation x
n+1
= f (x
n
) where f is the piecewise linear
function
f (x) =
_
_
_
−3x if x ≤ 100
0 if x > 100.
(a) Show that x
∗
= 0 is an unstable ﬁxed point.
(b) Despite the instability, show that x
∗
= 0 is globally attracting! That is,
regardless of the choice of initial condition x
0
, we have x
n
→ x
∗
as n →∞.
Hint: In fact, the iterates x
n
converge to x
∗
after ﬁnitely many iterations.
8. Sometimes, even the Schwarzian derivative test for stability (Theorem 7.3.12)
can fail to help us. Consider the difference equation
x
n+1
=
2
x
n
.
(a) Find all ﬁxed points of this equation, and show that they are non
hyperbolic.
(b) Show that Theorem 7.3.12 is inconclusive.
(c) Suppose we choose any nonzero initial condition x
0
, excluding the ﬁxed
points. Show that the solution of the difference equation will always
alternate between two distinct values.
(d) Explain why Part (c) allows us to conclude that the ﬁxed points are stable,
but not asymptotically stable.
9. Computing square roots by hand. Suppose k is a positive real number and
consider the difference equation
x
n+1
= x
n
−
x
2
n
−k
2x
n
.
216 how to control chaos
(a) Show that x
∗
=
√
k is a ﬁxed point and is locally asymptotically stable.
(b) The result of Part (a) implies that if we start with any initial condition
x
0
that is reasonably close to
√
k, then the sequence of iterates x
n
will
converge rapidly to
√
k. Try this with k = 2, using an initial guess of
x
0
= 1. Compute x
1
, x
2
, and x
3
. If you do this correctly, you will ﬁnd that
x
3
approximates
√
2 accurate to within 2.2 10
−6
.
10. Find all ﬁxed points of the system
x
n+1
= 3x
n
−x
n
y
n
y
n+1
= −2y
n
+ x
n
y
n
and classify the local stability of each ﬁxed point.
11. Consider the difference equation
x
n+1
= µ(x
n
−1)(2 −x
n
),
where µ is a positive parameter.
(a) Show that there are two ﬁxed points if µ < 3 −
√
8, zero ﬁxed points if
3 −
√
8 < µ < 3 +
√
8, and two ﬁxed points if µ > 3 +
√
8.
(b) Show that if µ = 6, then the smaller ﬁxed point is unstable and the larger
ﬁxed point is locally asymptotically stable.
(c) Show that if µ = 7, then both ﬁxed points are unstable. A perioddoubling
bifurcation occurs between µ = 6 and µ = 7. If x
∗
denotes the larger of
the two ﬁxed points, then f
/
(x
∗
) decreases below −1 for some µ between
6 and 7.
12. In this exercise, you will apply the tdas control algorithm to terminate a
period2 cycle in the discrete logistic equation.
(a) Using Equation (7.1) with µ = 3.2 and initial condition x
0
= 0.5, compute
x
1
, x
2
, . . . x
20
.
(b) To simulate tdas control, now we will suddenly switch from Equation (7.1)
to Equation (7.6). Let γ = 1.0. Using the value for x
20
that you computed
in Part (a) as an initial condition, generate the ﬁrst 20 iterates of Equa
tion (7.6). Describe what happens.
(c) Repeat Part (b) with γ = 0.3. Does tdas successfully stop the alternation?
Explain. (You may wish to refer to Figure 7.3.)
i ntroducti on to di fference equati ons 217
13. In this exercise, you will apply the tdas control algorithm to terminate a
period4 cycle in the discrete logistic equation.
(a) Using Equation (7.1) with µ = 3.5 and initial condition x
0
= 0.5, compute
x
1
, x
2
, . . . x
20
.
(b) To simulate tdas control, now we will suddenly switch from Equation (7.1)
to Equation (7.6). Let γ = 2.0. Using the value for x
20
that you computed
in Part (a) as an initial condition, generate the ﬁrst 20 iterates of Equa
tion (7.6). Describe what happens.
CHAPTER 8
Introduction to Partial Diﬀerential Equations
M
any of the natural phenomena that we wish to mathematically model
involve several independent variables. For example, the outdoor tem
perature T depends not only upon time, but also upon spatial location. If x and
y denote latitude and longitude and t denotes time, then the function T(x, y, t)
describes how temperature varies in space and time. Weather reports usually
render this function by using animations in which the variable t increases. In
stead of plotting T(x, y, t) as a surface in three dimensions for each ﬁxed t, the
maps are usually colorcoded, with red corresponding to high temperature and
blue corresponding to low temperature.
Mathematical models of phenomena incorporating several independent vari
ables frequently lead to equations involving partial derivatives. Usually, the
independent variables correspond to time and position. Before deﬁning what
we mean by a partial differential equation
1
, let us establish notation. If u is
a quantity that depends upon a single spatial variable (e.g., latitude) as well
as time, we will usually write u = u(x, t). Here x denotes the spatial variable
and t denotes time. When three spatial dimensions are involved, we will write
u = u(x, y, z, t). Instead of using the Leibniz notation for partial derivatives, we
use subscripts as follows:
u
x
=
∂u
∂x
u
t
=
∂u
∂t
u
xx
=
∂
2
u
∂x
2
u
tt
=
∂
2
u
∂t
2
u
xt
= (u
x
)
t
=
∂
∂t
_
∂u
∂x
_
=
∂
2
u
∂t∂x
.
1
All of our subsequent presentation is based heavily on the text of Strauss [10].
218
i ntroducti on to parti al di fferenti al equati ons 219
Unless otherwise stated, we will always assume that our functions u are suf
ﬁciently smooth to ensure that mixed partial derivatives are equal. That is,
u
xt
= u
tx
and u
xxt
= u
xtx
= u
txx
, and so on.
Roughly speaking, a partial differential equation is any equation involving
partial derivatives of some function u. With the above notational conventions in
mind, we state a more precise deﬁnition.
Deﬁnition 8.0.2. Suppose u = u(x, t) is a function of two variables. An equation
of the form
F(x, t, u, u
x
, u
t
, u
xx
, u
xt
, u
tt
, . . .) = 0
is called a partial differential equation (pde).
In this deﬁnition, it is understood that the function F has only ﬁnitely many
arguments. The deﬁnition is easily extended to allow for more than two in
dependent variables. The order of a pde is the order of the highest derivative
present in the equation. For example, the equation u
t
+ u
x
= 0 is a ﬁrstorder
pde, and the equation u
xxt
−(u
x
)
8
= 0 is a thirdorder pde. The most general
form of a ﬁrstorder pde with three independent variables t, x, and y would be
F(t, x, y, u, u
t
, u
x
, u
y
) = 0. Here are some wellknown examples of pdes.
The transport or advection equation: Let u = u(x, t). Then the equation u
t
+
cu
x
= 0 where c is a constant is called the simple advection equation. It can be
used to model the transport of a pollutant being carried (but not diffusing) in a
long, straight river with velocity c.
The heat or diffusion equation: Let u = u(x, t). Then the equation u
t
= κu
xx
where κ > 0 is a constant is called the simple diffusion equation. It can be used to
model the transfer of heat in a long, thin wire or diffusion of a dye in water.
The wave equation: Let u = u(x, t). Then the equation u
tt
−c
2
u
xx
= 0 where c
is a constant is called the wave equation. It can be used to model the displacement
of a plucked guitar string.
Burgers’ equation: Let u = u(x, t). Then the equation u
t
+ uu
x
= 0 is called
Burgers’ equation. It arises in the study of shock waves.
220
Laplace’s equation: Let u = u(x, y, z). Then the equation u
xx
+ u
yy
+ u
zz
= 0 is
called Laplace’s equation. It models steadystate heat distribution.
Kortewegde Vries equation: Let u = u(x, t). Then the thirdorder equation
u
t
+ u
xxx
−6uu
x
= 0 is called the Kortewegde Vries equation. It arises in the
context of modeling shallowwater surface waves.
Deﬁnition 8.0.3. A solution of the pde
F(x, t, u, u
x
, u
t
, u
xx
, u
xt
, u
tt
, . . .) = 0
is any function u(x, t) that satisﬁes the equation.
This deﬁnition, which is easily extended to allow for more than two indepen
dent variables, tacitly assumes the existence of partial derivatives of u. When we
use the word solution, we refer to what more advanced textbooks would call a
classical solution. It is possible to introduce the notion of weak solutions of a pde,
which does not require that the partial derivatives of u exist for all x and t.
Example 8.0.4. Consider the ﬁrstorder constantcoefﬁcient pde αu
x
+ βu
y
= 0.
We claim that u(x, y) = cos(βx − αy) is a solution of this pde. Taking the
ﬁrstorder partial derivatives,
∂
∂x
cos(βx −αy) = −β sin(βx −αy) and
∂
∂y
cos(βx −αy) = α sin(βx −αy).
Therefore,
αu
x
+ βu
y
= −αβ sin(βx −αy) + αβ sin(βx −αy) = 0,
which proves the claim. You can also show that the function u(x, y) = (βx −αy)
3
is a solution of the same pde. This may seem surprising, since the cosine function
and the cubic function are very different.
Example 8.0.5. For (x, y) ,= (0, 0), the function u(x, y) = ln(x
2
+y
2
) is a solution
of the twodimensional Laplace equation u
xx
+ u
yy
= 0. To see why, start by
taking the ﬁrstorder partial derivatives
u
x
=
2x
x
2
+ y
2
and u
y
=
2y
x
2
+ y
2
.
i ntroducti on to parti al di fferenti al equati ons 221
Using the quotient rule, the relevant secondorder partial derivatives are
u
xx
=
2(x
2
+ y
2
) −(2x)(2x)
(x
2
+ y
2
)
2
=
−2x
2
+2y
2
(x
2
+ y
2
)
2
,
u
yy
=
2(x
2
+ y
2
) −(2y)(2y)
(x
2
+ y
2
)
2
=
2x
2
−2y
2
(x
2
+ y
2
)
2
.
Adding these expressions, we see that u
xx
+ u
yy
= 0 as claimed.
In subsequent chapters, we will develop analytical techniques for determining
all solutions of certain ﬁrst and secondorder pdes. First, we introduce several
notions that can be used to classify pdes (e.g., linearity and homogeneity). By
simply classifying a pde, we can often determine whether (i) it is reasonable to
expect to solve the pde by hand, and (ii) if so, which solution techniques are
most likely to succeed.
8.1. Basic Classiﬁcation of Partial Diﬀerential Equations
In odes, we learn that linear equations are amenable to analytical techniques,
whereas nonlinear ones are usually intractable. The same principle holds for
pdes, and therefore it is useful to determine whether a pde is linear or nonlinear.
A brief review of some notions from linear algebra will assist in our deﬁnition of
linearity.
Linearity. When we think of functions, we usually have in mind a rule that
assigns numbers to numbers—for example, f : R →R. An operator is essentially
a special type of function that acts on a function to produce another function. One
example of an operator is the differentiation operator T =
d
dx
. If we feed T a
function of the variable x, it returns another function. For example, T(x
3
) = 3x
2
,
and T(e
x
) = e
x
. The (implied) domain of T is the set of all differentiable
functions of x.
Deﬁnition 8.1.1. An operator L is called linear if L(αu + βv) = αL(u) + βL(v)
for all functions u, v and all real scalars α, β. Equivalently, L is a linear operator
if both L(u + v) = L(u) + L(v) and L(cu) = cL(u) for all functions u and v and
scalars c.
222 basi c classi fi cati on of parti al di fferenti al equati ons
Example 8.1.2. Consider the set S of all functions f : R → R that are differen
tiable on the entire real line. The differential operator L =
d
dx
is a linear operator
on the set S. To see this, suppose u(x) and v(x) are in S and let α and β be any
real constants. Then
L(αu + βv) =
d
dx
(αu + βv) = α
d
dx
u + β
d
dx
v = αL(u) + βL(v).
Example 8.1.3. Consider the set S of all continuous functions on the closed
interval x ∈ [a, b]. For functions f (x) ∈ S, deﬁne the integral operator
I( f (x)) =
_
x
a
f (t) dt (a ≤ x ≤ b).
For example, if f (x) = cos(x), then I( f (x)) = sin x −sin a, and we see that I
transforms functions of x into new functions of x. We claim that I is a linear
operator. Indeed, suppose that u and v are functions in the set S, and let c
1
and
c
2
be any real numbers. Then
I(c
1
u + c
2
v) =
_
x
a
(c
1
u(t) + c
2
v(t)) dt = c
1
_
x
a
u(t) dt + c
2
_
x
a
v(t) dt
= c
1
I(u) + c
2
I(v).
Example 8.1.4. Let S be the set of all continuous functions from R into R. The
operator Φ deﬁned by the rule Φ(u) = u
2
is a nonlinear operator. To see
why, suppose u and v are functions from the set S. Notice that Φ(u + v) =
(u + v)
2
= u
2
+2uv + v
2
, whereas Φ(u) +Φ(v) = u
2
+ v
2
. In general, it is not
the case that Φ(u + v) = Φ(u) +Φ(v), and therefore Φ is a nonlinear operator.
Example 8.1.5. New linear operators can be formed by taking combinations of
other linear operators. For example, consider the set S of all functions u(x, t)
that are differentiable with respect to both x and t. The operator
L =
_
t
2
∂
∂x
+e
x
∂
∂t
_
is linear. Given a function u(x, t), the operator L acts on u according to the rule
L(u) =
_
t
2
∂
∂x
+e
x
∂
∂t
_
u = t
2
∂u
∂x
+e
x
∂u
∂t
.
i ntroducti on to parti al di fferenti al equati ons 223
Now suppose that u and v are functions from the set S. Then
L(u + v) =
_
t
2
∂
∂x
+e
x
∂
∂t
_
(u + v) = t
2
∂
∂x
(u + v) +e
x
∂
∂t
(u + v)
= t
2
_
∂u
∂x
+
∂v
∂x
_
+e
x
_
∂u
∂t
+
∂v
∂t
_
.
Comparing the latter expression to
L(u) + L(v) =
_
t
2
∂
∂x
+e
x
∂
∂t
_
u +
_
t
2
∂
∂x
+e
x
∂
∂t
_
v
= t
2
∂u
∂x
+e
x
∂u
∂t
+ t
2
∂v
∂x
+e
x
∂v
∂t
,
we see that L(u + v) = L(u) + L(v). It is also straightforward to show that if
u ∈ S and c is any constant, then L(cu) = cL(u). Therefore, the operator L is
linear.
The latter example motivates our deﬁnition of linearity for pdes. In what fol
lows, we restrict ourselves to functions u = u(x, t) of two independent variables,
although these concepts are readily extended to more general cases. Moreover,
we assume that L is a linear partial differential operator—a linear operator which
incorporates partial derivatives with respect to at least one of the independent
variables.
Deﬁnition 8.1.6. A pde is called linear if it can be written in the form L(u) =
f (x, t), where L is a linear operator of the sort described in the preceding
paragraph. The function f (x, t) can be any function of the two independent
variables.
Example 8.1.7. Let κ be a positive constant. The heat equation
∂u
∂t
= κ
∂
2
u
∂x
2
is a linear pde. Writing the equation in the form
∂u
∂t
−κ
∂
2
u
∂x
2
= 0 suggests that we
deﬁne the operator
L =
_
∂
∂t
−κ
∂
2
∂x
2
_
and the function f (x, t) = 0. The pde takes the form L(u) = f (x, t), and you can
show that L is a linear operator. It follows that the heat equation is linear.
224 basi c classi fi cati on of parti al di fferenti al equati ons
Example 8.1.8. The pde
_
1 + x
2
+ t
2
∂
2
u
∂x
2
+ 3 sin(xt)
∂
2
u
∂t
2
−8 ln(1 + x
2
+ t
4
) = 0
is linear. If we deﬁne the operator
L =
_
_
1 + x
2
+ t
2
_
∂
2
∂x
2
+ (3 sin xt)
∂
2
∂t
2
and the function f (x, t) = 8 ln(1 + x
2
+ t
4
), then the pde takes the form
L(u) = f (x, t). To demonstrate that L is a linear operator, use computations
similar to those in the preceding example.
Example 8.1.9. Burgers’ equation u
t
+ uu
x
= 0 is a nonlinear pde. If we deﬁne
the operator N according to the rule
N(u) =
_
∂
∂t
+ u
∂
∂x
_
u = u
t
+ uu
x
,
then Burgers’ equation takes the form N(u) = 0. To see that the operator N is
nonlinear, suppose that u and v are functions. Then
N(u + v) =
_
∂
∂t
+ (u + v)
∂
∂x
_
(u + v) =
∂
∂t
(u + v) + (u + v)
∂
∂x
(u + v)
=
∂u
∂t
+
∂v
∂t
+ u
∂u
∂x
+ u
∂v
∂x
+ v
∂u
∂x
+ v
∂v
∂x
.
Comparing the latter expression with
N(u) + N(v) =
_
∂
∂t
+ u
∂
∂x
_
u +
_
∂
∂t
+ v
∂
∂x
_
v =
∂u
∂t
+ u
∂u
∂x
+
∂v
∂t
+ v
∂v
∂x
,
in general it is not the case that N(u + v) = N(u) + N(v). Therefore the pde is
nonlinear.
Homogeneity. One further way of classifying linear pdes is provided by the
following deﬁnition. As before, we assume that u = u(x, t) is a function of two
independent variables and that L is a linear partial differential operator.
Deﬁnition 8.1.10. A linear pde L(u) = f (x, t) is called homogeneous if f (x, t) = 0.
Otherwise, the pde is called inhomogeneous.
i ntroducti on to parti al di fferenti al equati ons 225
Example 8.1.11. The equation
∂u
∂t
−
∂
2
u
∂x
2
−8 = 0
is inhomogeneous. In order to write this pde in the form L(u) = f (x, t), we
would deﬁne
L =
_
∂
∂t
−
∂
2
∂x
2
_
and f (x, t) = 8. Since f ,= 0, the pde is inhomogeneous.
As in odes, the combination of linearity and homogeneity is very advantageous,
because it allows us to construct new solutions by the “superposition principle”.
Theorem 8.1.12 (Superposition Principle). Suppose L(u) = 0 is a linear, homo
geneous pde with particular solutions u
1
, u
2
, . . . u
n
. Then any linear combination
c
1
u
1
+ c
2
u
2
+ · · · + c
n
u
n
is also a solution, for any choice of constants c
1
, c
2
, . . . c
n
.
Proof. Since u
1
, u
2
, . . . u
n
are solutions of the pde, it follows that L(u
k
) = 0 for
each k = 1, 2, . . . n. Now let u = c
1
u
1
+ c
2
u
2
+ · · · + c
n
u
n
. Then by linearity of L,
L(u) = L
_
n
∑
k=1
c
k
u
k
_
=
n
∑
k=1
L (c
k
u
k
) =
n
∑
k=1
c
k
L(u
k
) = 0.
Therefore, u is also a solution of the pde.
Example 8.1.13. In a previous example, we noted that u
1
(x, y) = cos(βx −αy)
and u
2
(x, y) = (βx −αy)
3
are solutions of the pde αu
x
+ βu
y
= 0, where α and β
are constants. Since the pde is both linear and homogeneous, the Superposition
Principle (Theorem 8.1.12) ensures that
3u
1
−5u
2
= 3 cos(βx −αy) −5(βx −αy)
3
is also a solution of the pde.
Secondorder constantcoefﬁcient partial differential equations. Three of the
most important examples of pdes that we will analyze in subsequent chapters are
the heat equation u
t
−u
xx
= 0, the wave equation u
tt
−u
xx
= 0, and the Laplace
226 basi c classi fi cati on of parti al di fferenti al equati ons
equation u
xx
+ u
tt
= 0. In some sense, every secondorder constantcoefﬁcient
pde will behave like one of these three special equations. (This can be shown via
appropriate changes of variables; for an explicit example, see Strauss [10].) Since
the heat, wave, and Laplace equations have very different solutions, classifying
secondorder pdes as either “heatlike”, “wavelike”, or “Laplacelike” can be
very illuminating.
Consider the general, secondorder, homogeneous, constantcoefﬁcient pde
Au
xx
+ Bu
xt
+ Cu
tt
+ Du
x
+ Eu
t
+ Fu = 0, (8.1)
where A, B, C, D, E, and F are constants.
Deﬁnition 8.1.14. Equation (8.1) is called
Elliptic if B
2
−4AC < 0,
Hyperbolic if B
2
−4AC > 0,
Parabolic if B
2
−4AC = 0 and A, B, and C are not all zero.
Example 8.1.15. Consider the wave equation u
xx
−u
tt
= 0. Using the notation
of Deﬁnition 8.1.14, we have A = 1, C = −1, and B = D = E = F = 0. Since
B
2
−4AC = 4 > 0, we see that the wave equation is hyperbolic. For the heat
equation u
t
−u
xx
= 0, we have A = −1 and B = C = 0, from which it follows
that B
2
−4AC = 0. Thus, the heat equation is parabolic. Finally, the Laplace
equation u
xx
+u
tt
= 0 satisﬁes A = C = 1 and B = 0, implying that this equation
is elliptic.
To reinforce this terminology, it is useful to note the parallels between these
pdes and corresponding algebraic equations. For example, we can associate the
wave equation u
xx
−u
tt
= 0 with the algebraic equation x
2
−t
2
= 0, the graph
of which is a hyperbola in the xtplane. Similarly, the heat equation u
t
−u
xx
= 0
can be associated with the algebraic equation t − x
2
= 0, the graph of which
is a parabola in the xtplane. Generally, solutions of parabolic pdes tend to
behave similarly to those of the heat equation. Linear, hyperbolic pdes tend to be
“wavelike”, and elliptic pdes are “Laplacelike” in terms of the behavior of their
solutions.
i ntroducti on to parti al di fferenti al equati ons 227
8.2. Solutions of Partial Diﬀerential Equations
Finding all solutions of a pde is considerably more challenging than solving and
algebraic or ordinary differential equation. Solutions of algebraic equations such
as x
2
−6x +5 = 0 are numbers: x = 1 and x = 5. Solutions of odes are functions
of the independent variable. For example, the general solution of the equation
d
2
y
dx
2
−6
dy
dx
+5y = 0
is y(x) = C
1
e
x
+ C
2
e
5x
, where C
1
and C
2
are arbitrary constants. We know that
the general solution of a linear m
th
order ode will contain m arbitrary constants.
What do general solutions of pdes look like?
Example 8.2.1. Find all functions u(x, t) satisfying the pde u
x
= 0. Solution: We
are given the x derivative of u, and we need only integrate with respect to x in
order to solve for u. Our temptation may be to write u(x, t) = constant, but this
is incorrect—after all, there is a second independent variable, t. Thus, instead
of including an arbitrary integration constant when we integrate with respect to
x, we must include an arbitrary function of t. The general solution of this pde is
u(x, t) = f (t), where f (t) is any arbitrary function of t.
Example 8.2.2. Solve the pde u
xt
= 3. Solution: This secondorder pde can also
be solved by integration: since (u
x
)
t
= 3, we integrate with respect to t, treating
x as constant. Doing so reveals that u
x
= 3t + f (x), where f (x) is an arbitrary
function of x. Next, we integrate with respect to x, treating t as a constant.
We ﬁnd that u(x, t) = 3xt + F(x) + g(t), where F(x) =
_
f (x) dx represents an
antiderivative of f and g(t) is an arbitrary function of t. Since f (x) was arbitrary,
so is F(x). It follows that any function of the form u(x, t) = 3xt + F(x) + g(t)
will satisfy our original pde, and there are no other solutions.
Observation. Whereas general solutions of odes contain arbitrary constants, the
general solutions of pdes contain arbitrary functions of the independent variables.
This feature of pdes opens the door to having tons of solutions, and later we will
devote considerable effort toward singling out particular (physically relevant)
solutions. Doing so will require us to invoke physical intuition regarding the
correct number (and types) of initial and boundary conditions to impose.
228 i ni ti al condi ti ons and boundary condi ti ons
Example 8.2.3. Solve the linear secondorder pde u
tt
+ u
t
= 0 by ﬁnding all
functions u(x, t) that satisfy it. Solution: If we were not told that there is a
second independent variable x, we would have no way of distinguishing this pde
from an ode. Ignoring the variable x for the moment, consider the analogous
ode u
//
(t) + u
/
(t) = 0. Using techniques from your ﬁrst course in odes, you
could solve the associated characteristic equation and ﬁnd the general solution
u(t) = C
1
+ C
2
e
−t
, where C
1
and C
2
are arbitrary constants. Returning to our
original pde, we obtain the general solution by replacing the constants C
1
and
C
2
by arbitrary functions of x: that is, u(x, t) = f (x) + e
−t
g(x) is the general
solution of the original pde.
Another way to attack this problem is to note that we could have integrated
once with respect to t to reduce the pde from secondorder to ﬁrstorder:
_
u
tt
+ u
t
dt = 0.
Integration yields u
t
+u− f (x) = 0, where the minus sign in front of the arbitrary
function f (x) is merely included for convenience. The resulting inhomogeneous
ﬁrstorder equation u
t
+ u = f (x) can be solved by the variation of parameters
technique (see Theorem 2.18). Multiplying both sides of this ﬁrstorder equation
by the integrating factor e
t
, we have e
t
(u
t
+ u) = e
t
f (x). Equivalently,
∂
∂t
_
e
t
u
_
= e
t
f (x).
Integration with respect to t yields
e
t
u =
_
e
t
f (x) dt = e
t
f (x) + g(x),
where g(x) is an arbitrary function of x. Finally, multiplying both sides by e
−t
yields the general solution of the original pde, u(x, t) = f (x) +e
−t
g(x).
8.3. Initial Conditions and Boundary Conditions
In our study of odes, we learned that it is important to single out speciﬁc
solutions that are of particular physical relevance. For odes, we typically specify
initial conditions (or sometimes boundary conditions) in such a way that we may
solve for any arbitrary constants that appear in the general solution. The general
i ntroducti on to parti al di fferenti al equati ons 229
solutions of pdes contain arbitrary functions of the independent variables, making
it more difﬁcult to single out particular solutions. For example, suppose that
a pollutant
2
is carried by water moving with constant speed c through a long,
“onedimensional” pipe. Letting u(x, t) represent the concentration of pollutant
at position x and time t, the advection equation u
t
+ cu
x
= 0 can be used to
model the spatiotemporal dynamics of u. Since we would expect the general
solution of this ﬁrstorder pde to contain at least one arbitrary function of the
variables x and t, we should impose at least one auxiliary condition on u(x, t) in
order to select a particular solution.
Initial Conditions. Perhaps the most intuitive way to prescribe an auxiliary
condition is to impose an initial condition—i.e., give a formula for u(x, t) at
some ﬁxed time t = t
0
. In our above example of a transported pollutant, the
initial condition should describe the concentration of pollutant at all spatial
locations at time t
0
. Mathematically, such an initial condition would have the
form u(x, t
0
) = φ(x), where φ is a function only of the spatial variable x.
Example 8.3.1. Soon, we will know how to show that the general solution of
the advection equation u
t
+ 8u
x
= 0 on the domain −∞ < x < ∞ is given by
u(x, t) = f (x −8t), where f is any differentiable function of a single variable. To
verify that any such function satisﬁes the advection equation, we calculate the ﬁrst
partial derivatives of u(x, t) = f (x −8t) using the chain rule: u
t
= −8f
/
(x −8t)
and u
x
= f
/
(x −8t). Hence, u
t
+8u
x
= 0, as claimed. Now suppose we impose
an initial condition at time t = 0 by requiring that u(x, 0) = x
2
+cos x. Setting
t = 0 in our general solution u(x, t) = f (x −8t) implies that u(x, 0) = f (x),
from which it follows (by the initial condition) that f (x) = x
2
+ cos x. Finally,
we see that the only function satisfying both the pde and the initial condition is
u(x, t) = f (x −8t) = (x −8t)
2
+cos(x −8t).
In the above example, observe that only one auxiliary condition was needed
in order to isolate a particular solution. It is also worth noting that (i) the pde
was ﬁrstorder in both space and time and (ii) the spatial domain −∞ < x < ∞
consisted of all real numbers. As we shall see, imposing initial and boundary
2
Assume that the pollutant is simply carried by the current and does not diffuse into the water.
In other words, think of small, plastic beads as opposed to a chemical dye.
230 i ni ti al condi ti ons and boundary condi ti ons
conditions requires more care (and physical intuition) when dealing with higher
order pdes and/or pdes in which the spatial domain is ﬁnite.
Example 8.3.2. Consider the wave equation u
tt
−4u
xx
= 0 on the domain −∞ <
x < ∞. Later, we will learn how to show that the general solution of this pde
is given by u(x, t) = f (x + 2t) + g(x −2t), where f and g are arbitrary [twice
differentiable] functions of a single variable. Suppose that we impose two initial
conditions: u(x, 0) = cos(x) and u
t
(x, 0) = sin(x). From the general solution of
the pde, we use the chain rule to calculate
u
t
(x, t) = 2f
/
(x +2t) −2g
/
(x −2t).
Setting t = 0 in our expressions for u(x, t) and u
t
(x, t) allows us to use our initial
conditions:
u(x, 0) = f (x) + g(x) = cos x
u
t
(x, 0) = 2f
/
(x) −2g
/
(x) = sin x.
Taking the derivative of the equation f (x) + g(x) = cos x yields f
/
(x) + g
/
(x) =
−sin x which, combined with the other equation 2f
/
(x) − 2g
/
(x) = sin x al
lows us to algebraically solve for f
/
(x) and g
/
(x). Speciﬁcally, we ﬁnd that
f
/
(x) = −
1
4
sin x and g
/
(x) = −
3
4
sin x. Integrating with respect to x, we have
f (x) =
1
4
cos x + C
1
and g(x) =
3
4
cos x + C
2
, where C
1
and C
2
are constants of
integration. However, we may drop these integration constants because (i) they
must sum to zero in order for the ﬁrst initial condition to be satisﬁed and (ii) the
overall solution involves a sum of f and g. It follows that the only solution of
the wave equation satisfying both initial conditions is
u(x, t) =
1
4
cos(x +2t) +
3
4
cos(x −2t).
When we discuss the physical interpretation of the wave equation on the domain
−∞ < x < ∞, we will see that prescribing two initial conditions u(x, 0) and
u
t
(x, 0) is analogous to stating an initial position and velocity.
When the spatial domain is inﬁnite and no boundary conditions are assigned,
a pde together with its initial conditions is sometimes referred to as the Cauchy
problem, which is basically another way of saying “initial value problem”. In the
preceding two examples, the ﬁrst shows a solution of the Cauchy problem for
i ntroducti on to parti al di fferenti al equati ons 231
the advection equation, and the second shows a solution of the Cauchy problem
for the wave equation.
Boundary Conditions. In contrast with the preceding two examples, for practi
cal purposes we are usually interested in solving pdes over ﬁnite spatial domains.
Depending upon what sorts of phenomena our pdes are meant to model, we may
have special information about the physical state of our system at the boundaries
of the spatial domain. When imposing a boundary condition, we must describe
the physical state of our system at a speciﬁc spatial location as time t varies.
Contrast this with how initial conditions are prescribed. Boundary conditions
“freeze” the spatial variable x and describe the physical state of the system at that
location as t varies. Initial conditions “freeze” the time variable t and describe
the physical state of the system at all spatial locations x.
To give an intuitive illustration of how boundary conditions arise, let us
consider the heat equation u
t
− u
xx
= 0 in one spatial dimension. This pde
models how heat diffuses in a long, thin insulated
3
wire. Letting L denote the
length of the wire, we may assume that our spatial domain is given by the
interval 0 ≤ x ≤ L. The boundary of our spatial domain consists of two points:
x = 0 and x = L. One way to impose boundary conditions at would be to simply
give formulas for u at these two points—i.e., specify u(0, t) and u(L, t) for all
time t ≥ 0. A boundary condition which speciﬁes the exact value of u at the
boundary is called a Dirichlet condition. For example, if we dip the x = L end of
the wire in a bath of boiling water of constant temperature 100 degrees Celsius,
this imposes the Dirichlet boundary condition u(L, t) = 100 for all t ≥ 0. If the
other end of the wire is placed in contact with a block of ice of temperature 0
degrees Celsius, that would impose another Dirichlet condition u(0, t) = 0 for all
t ≥ 0. Dirichlet boundary conditions can also allow us to vary the temperature
at the boundary. For example, u(L, t) = 80 +20 cos(t) would simulate the effects
of varying the temperature between 60 and 100 degrees at the x = L end of the
wire.
Another way to impose a boundary condition is to describe the temperature
gradient at the boundary of the spatial domain. Suppose that we insulate the
x = 0 end of the wire so that no heat can escape (or enter) the wire at the
point. Then there can be no spatial gradient of heat at x = 0, implying that
3
The wire’s insulation prevents heat from diffusing outward from the wire so that it is only
conducted longitudinally within the wire.
232 i ni ti al condi ti ons and boundary condi ti ons
u
x
(0, t) = 0 for all t ≥ 0. The boundary condition u
x
(0, t) = 0 is an example of a
Neumann condition. In general, a Neumann condition is a boundary condition
that describes the outward normal derivative of u along the boundary. If our
spatial domain is given by an interval 0 ≤ x ≤ L in one dimension, Neumann
conditions specify u
x
(0, t) and u
x
(L, t) for all time t. For example, the Neumann
condition u
x
(L, t) = t would indicate that the spatial temperature gradient at
x = L increases linearly in time, with higher temperature outside the wire
(x > L) than inside the wire x < L. Assuming that heat ﬂows in the direction
opposite the heat gradient (i.e., from regions of higher temperature towards
regions of lower temperature), the steepening gradient would cause heat to ﬂow
from right to left (into the wire) at the boundary x = L.
In more than one spatial dimension, the boundary of our domain is typically
a curve, surface, or hypersurface. For example, Figure 8.1 shows a domain Ω
in two space dimensions. The boundary of Ω, sometimes denoted as ∂Ω, is a
closed curve. If x and y are the spatial variables, then a Dirichlet condition would
give a formula for the value of u(x, y, t) for all points (x, y) ∈ ∂Ω for all time t.
A Neumann condition would be used to describe the net outward ﬂow across
the boundary ∂Ω for all time t. More speciﬁcally, let ∇u = (u
x
, u
y
) denote the
gradient of u with respect to its spatial variables, and let n denote an outward
unit normal vector on ∂Ω (see ﬁgure). Neumann boundary conditions would
provide a formula for ∇u • n for all points (x, y) ∈ ∂Ω. Assuming that ∇u is not
the zero vector, we know that ∇u points in the direction in which u increases
most rapidly. The dot product ∇u • n indicates whether the gradient vector ∇u
points outward from Ω (if ∇u • n > 0), inward (if ∇u • n < 0), or tangent to the
boundary (if ∇u • n = 0). In the case of the heat/diffusion equation, heat diffuses
from regions of high temperature towards regions of lower temperature, ﬂowing
in the direction opposite the heat gradient. If u(x, y, t) denotes temperature at
position (x, y) ∈ Ω at time t and we ﬁnd that ∇u • n > 0 at all points on ∂Ω,
then heat would ﬂow inward into our domain Ω. If we insulate the boundary to
prevent heat ﬂow across ∂Ω, then we impose the Neumann condition ∇u • n = 0
at all points on ∂Ω.
Wellposed problems. Knowing how to prescribe initial and boundary condi
tions in such a way that we single out precisely one physically relevant solution
of our pdes is very important. Ideally, we always hope to work with wellposed
problems—those with exactly one solution and for which small changes in initial
i ntroducti on to parti al di fferenti al equati ons 233
Ω
n
n
n
Figure 8.1. A spatial domain Ω in two dimensions. Several outward
unit normal vectors n are shown.
and boundary conditions do not elicit major changes in the behavior of the
solution. As we learn more about the physical interpretation of certain pdes, we
will learn how to impose initial and boundary conditions in such a way that our
problems are wellposed.
8.4. Visualizing Solutions of Partial Diﬀerential Equations
Visualizing the behavior of solutions for systems of ordinary differential equa
tions is rather straightforward. If t is the independent variable and
x
1
(t), x
2
(t), . . . x
n
(t)
are the dependent variables, then one way to graphically render the solution is
to plot each dependent variable versus t. Now suppose that u(x, t) represents a
solution of a pde with exactly two independent variables, x and t. At ﬁrst glance,
you may be tempted to plot u as a function of both x and t, but this would
require a threedimensional plot in which u is plotted as a surface. Although this
is certainly a valid way of presenting the solution, it is often more convenient
to provide twodimensional plots that illustrate u versus x at several different
times t
1
, t
2
, . . . t
n
. This effectively provides us with frames of an “animation”,
illustrating the spatial distribution of u as we advance forward in time.
Example 8.4.1. Consider the pde u
t
+u
x
= 0 on the spatial domain −∞ < x < ∞,
and with the initial condition u(x, 0) = e
−x
2
. In the next Chapter, you will learn
234 vi suali zi ng soluti ons of parti al di fferenti al equati ons
6 12
x
u
(
x
,
t
)
t = 3 t = 6 t = 9 t = 0
0 3 9
Figure 8.2. Left panel: A surface plot of u(x, t) = e
−(x−t)
2
versus both
x and t. Right panel: A plot of u(x, t
∗
) versus x at four different choices
of t
∗
, namely 0, 3, 6, and 9. Note that the initial spatial distribution of
u at t = 0 is transported from left to right with constant speed 1.0 as t
increases.
to show that the solution of this initial value problem is given by
u(x, t) = e
−(x−t)
2
.
Two ways of rendering this solution are illustrated in Figure 8.2. The left panel
shows a plot of the function u(x, t) as a surface in three dimensions. The right
panel gives four frames of an “animation”, showing that our initial distribution
of u is transported from left to right with constant speed as t increases. The
four traces of u versus x suggest that the speed is 1.0 spatial units per unit
time. To further convince ourselves that the propagation speed is 1.0, we can
use an analytical approach to track the movement of the peaks. Setting u = 1
in our formula for u(x, t) yields 1 = e
−(x−t)
2
, and taking logarithms reveals that
(x −t)
2
= 0. To calculate the propagation velocity
dx
dt
, differentiate the latter
expression (implicitly) with respect to t, and use algebra to show that
dx
dt
= 1.
Exercises
1. Let S denote the set of functions of two variables (x and y) whose partial
derivatives of all orders exist. Which of the following operators T are linear
on the set S? Show computations to support your answers.
i ntroducti on to parti al di fferenti al equati ons 235
(a) T(u) = u
x
u
y
(b) T(u) = yu
x
+ xu
y
(c) T(u) = 1
(d) T(u) = [arctan(x)]u
xxx
(e) T(u) =
_
u
xx
+ u
yy
(f) T(u) = u
x
+ u
y
+ x.
2. Determine whether or not each of the following pdes is linear or nonlinear.
Show computations to support your answers.
u
x
+ uu
y
= 0
u
tt
−u
xx
+ u
3
= 0
u
tt
+ u
xxxx
= cos(xt).
3. For each of the following equations, state the order and whether it is linear
or nonlinear. If an equation is linear, state whether it is homogeneous or
inhomogeneous.
(a) u
t
−4u
xx
−ln(1 + x
2
) = 0
(b) u
tt
−u
xx
+ xu = 0
(c) u
t
+ u
xxx
−6uu
x
= 0
(d) u
2
x
−u
2
y
= 0
(e) u
t
+ u
xx
+
u
x
2
= 0
(f) u + u
xy
+
√
1 + u
2
= 0
(g) u
x
+e
−xy
u
y
+2 = 0
(h) u
x
u
y
−1 = 0
4. Let κ > 0 be a constant. Show that u(x, t) = A+ Bx +2κCt +Cx
2
is a solution
of the pde u
t
= κu
xx
for any choice of constants A, B, and C.
5. Suppose that f and g are arbitrary differentiable functions of a single variable,
and that c is a constant. Show that u(x, t) = f (x + ct) + g(x −ct) is a solution
of the pde u
tt
−c
2
u
xx
= 0.
6. Show that u(x, y) = cos(x
2
+ y
3
) is a solution of the pde u
xy
+6xy
2
u = 0.
7. Solve the pde u
xt
= 6xt
2
.
8. Find a constant solution of the initial value problem
dy
dx
= 3y
2/3
, y(0) = 0.
Then, use separation of variables to ﬁnd a nonconstant solution. Conclude
that this problem is not wellposed since it violates uniqueness.
9. Is the boundary value problem
dy
dx
+ y = 0, y(0) = 1, y(2) =
1
2
a wellposed problem? If so, ﬁnd the solution. If not, explain why not.
CHAPTER 9
Linear, FirstOrder Partial Diﬀerential Equations
I
n this chapter, we will discuss the ﬁrst of several special classes of pdes that
can be solve via analytical techniques. In particular, we will investigate linear,
ﬁrstorder pdes
a(x, t)
∂u
∂t
+ b(x, t)
∂u
∂x
+ f (x, t)u = g(x, t), (9.1)
where u = u(x, t) is our dependent variable, and the functions a, b, f , and g are
given. Our goal is to develop a systematic method for determining all functions
u that satisfy the pde. A little geometric intuition will help use devise a rather
useful technique for solving such equations, and for that reason we will review a
few notions from multivariate calculus.
Suppose that f (x, y) is a differentiable function of two variables, and let
v = (v
1
, v
2
) denote a unit vector in R
2
.
Deﬁnition 9.0.2. The directional derivative of f in the direction of the vector v is
given by
lim
h→0
f (x + v
1
h, y + v
2
h) − f (x, y)
h
,
provided that this limit exists.
It is straightforward to prove (see your multivariate calculus text) that this
limit is equal to the dot product
∇f (x, y) • v,
thereby providing us with an equivalent (and computationally convenient) def
inition of the directional derivative. The directional derivative measures the
236
li near, fi rst order parti al di fferenti al equati ons 237
instantaneous rate of change of f in the direction of the vector v. If we choose
the unit vector v = (1, 0), which points in the direction of the positive xaxis,
then the directional derivative is simply
∇f (x, y) • v =
_
∂ f
∂x
,
∂ f
∂y
_
• (1, 0) =
∂ f
∂x
,
the partial derivative of f with respect to x. Similarly, the directional derivative in
the direction of v = (0, 1) is given by ∂ f /∂y. We now use the notion of directional
derivatives to give a quick, clean solution of a special case of Equation (9.1).
A linear, homogeneous, constantcoefﬁcient equation. Consider the pde
αu
t
+ βu
x
= 0, (9.2)
where α and β are nonzero constants. Our goal is to determine all functions
u(x, t) that satisfy this equation. Observe that since the gradient of u is given by
∇u = (u
x
, u
t
), the pde (9.2) is equivalent to the equation
∇u(x, t) • (β, α) = 0.
In other words, the solutions of the pde are precisely those functions u(x, t) whose
directional derivative in the direction of the vector (β, α) is 0. Geometrically, this
implies that u(x, t) must remain constant as we move along any line parallel to
the vector (β, α). These lines have slope α/β in the xt plane, and the general
equation for such lines is t = (α/β)x + constant. Equivalently, u(x, t) must
remain constant along any line of the form αx − βt = C, where C is an arbitrary
constant. Since u remains constant along each such line, u depends only on
the value of the constant C. In other words, u depends only upon the quantity
αx − βt, but otherwise has no restrictions at all. It follows that the general
solution of Equation (9.2) is
u(x, t) = f (αx − βt), (9.3)
where f is any (differentiable) function of a single variable.
238
Deﬁnition 9.0.3. In the above example, the “geometric” technique used to solve
the pde is called the method of characteristics. The lines αx − βt = constant in the
above example are called characteristic curves associated with the pde.
By differentiation, it is easy to verify that any function of the form (9.3) certainly
satisﬁes the pde (9.2). The method of characteristics demonstrated that there are
no other solutions of the pde. Later, we will give more examples illustrating the
use of the method of characteristics.
Example 9.0.4. To check that u(x, t) = cos(αx −βt) is a solution of the pde (9.2),
we compute the partial derivatives u
t
= β sin(αx −βt) and u
x
= −α sin(αx −βt).
From here it is evident that αu
t
+ βu
x
= 0, as required.
Example 9.0.5. Solve the pde 2u
t
+ u
x
= 0 with initial condition u(x, 0) =
1
1+x
2
,
and sketch u versus x for several choices of time t. Solution: This pde is a special
case of Equation (9.2), so we may quote (9.3) to see that the general solution is
u(x, t) = f (2x −t), where f is an arbitrary differentiable function of a single
variable. To use the initial condition, we set t = 0 to ﬁnd that
f (2x) =
1
1 + x
2
.
Dividing x by 2 in the preceding equation leads us to
f (x) =
1
1 + (x/2)
2
.
Combining the general solution u(x, t) = f (2x −t) with our formula for f , the
overall solution of the initial value problem is
u(x, t) =
1
1 +
_
2x−t
2
_
2
=
1
1 +
_
x −
t
2
_
2
.
To visualize the solution, it is useful to plot u(x, t) versus x for several choices
of time t. Figure 9.1 shows a plot of u versus x at t = 0 (the initial condition)
as well as t = 10 and t = 20. Visual inspection suggests a constant propagation
speed of 0.5 units, and straightforward calculus reveals that this is, indeed, the
case.
The preceding example gives insight into the dynamics of the constant
coefﬁcient pde (9.2). Namely, the initial (t = 0) spatial distribution of u is
li near, fi rst order parti al di fferenti al equati ons 239
t = 0 t = 10 t = 20
u
(
x
,
t
)
0 5
10
x
Figure 9.1. A plot of u(x, t) versus x for t = 0, t = 10, and t = 20. The
three frames of this “animation” suggest that the initial distribution of
u is transported from left to right with speed 0.5 units.
transported with constant speed while retaining the same spatial proﬁle. We now
give an illustration of how this pde arises naturally in the context of modeling
transport.
9.1. Derivation and Solution of the Transport Equation
Assuming α ,= 0, we may divide Equation (9.2) by α to obtain
u
t
+ cu
x
= 0, (9.4)
where c = β/α is a constant. We will refer to this pde as the transport equation or
the advection equation.
Derivation. A physical interpretation of the transport equation can be given as
follows. Suppose that a ﬂuid moves with constant speed c in a long, thin pipe.
Assume that a pollutant is suspended in the water, and is simply carried by the
ﬂuid without diffusing. Moreover, we assume that the pipe can be treated as
onedimensional and that there is no drag/friction between the ﬂuid and the
walls of the pipe. If x represents our position along the length of the pipe, we
will let u(x, t) denote the concentration (mass per unit length) of pollutant at
position x and time t. The mass m of pollutant in an interval [0, x] at time t is
240 deri vati on and soluti on of the transport equati on
obtained by integrating concentration with respect to the length variable:
m =
_
x
0
u(s, t) ds.
If we advance ∆t units forward in time, the mass that was in the interval [0, x]
at time t has moved c∆t units down the pipe. In other words, that same mass
would be in the interval [c∆t, x + c∆t] at time t +∆t, implying that
m =
_
x+c∆t
c∆t
u(s, t +∆t) ds.
Equating our two expressions for mass,
_
x
0
u(s, t) ds =
_
x+c∆t
c∆t
u(s, t +∆t) ds.
Differentiating with respect to x and using the Fundamental Theorem of Calculus,
we ﬁnd that
u(x, t) = u(x + c∆t, t +∆t)
for all choices of ∆t. Equivalently, u(x + c∆t, t +∆t) −u(x, t) = 0 and, dividing
by ∆t, we have
u(x + c∆t, t +∆t) −u(x, t)
∆t
= 0.
Letting ∆t → 0,
lim
∆t→0
u(x + c∆t, t +∆t) −u(x, t)
∆t
= 0.
Referring to Deﬁnition 9.0.2, the preceding equation asserts that the directional
derivative in the direction of the vector (c, 1) is zero. Equivalently, (c, 1) •
∇u(x, t) = 0, which is identical to Equation (9.4).
The general solution of the transport equation is readily obtained from Equa
tion (9.3), namely u(x, t) = g(x − ct) where g is an arbitrary differentiable
function. The Cauchy problem (i.e., initial value problem on an inﬁnite spatial
domain) for the transport equation is given by
u
t
+ cu
x
= 0, (−∞ < x < ∞ and t > 0)
u(x, 0) = f (x), (−∞ < x < ∞ and t = 0).
li near, fi rst order parti al di fferenti al equati ons 241
Setting t = 0 in the general solution u(x, t) = g(x −ct) reveals that f (x) = g(x).
Therefore, the solution of the Cauchy problem for the transport equation is
u(x, t) = f (x −ct). Note that at t = 0, the spatial concentration proﬁle of the
pollutant is described by the graph of f (x). At t = 1, we have u(x, t) = f (x −c),
implying that the initial distribution of pollutant has been translated horizontally
by c units. If c > 0, the movement is from lefttoright, and if c < 0 then the
movement is from righttoleft.
The transport equation can be generalized to higherdimensional spatial do
mains. Let x = (x
1
, x
2
, . . . x
n
) ∈ R
n
and suppose that b = (b
1
, b
2
, . . . b
n
) is a
constant vector in R
n
. The transport equation is given by
u
t
+b • ∇u = 0,
where the gradient of u = u(x; t) is taken with respect to the spatial variables
only. That is,
u
t
+ b
1
∂u
∂x
1
+ b
2
∂u
∂x
2
+ · · · + b
n
∂u
∂x
n
= 0.
The general solution of this pde is given by
u(x; t) = u(x
1
, x
2
, . . . x
n
, t) = g(x −tb) = g(x
1
−tb
1
, x
2
−tb
2
, . . . , x
n
−tb
n
),
where g : R
n
→ R is an arbitrary differentiable function. It follows that the
solution of the Cauchy problem
u
t
+b • ∇u = 0, (x ∈ R
n
and t > 0)
u(x, 0) = f (x), (x ∈ R
n
and t = 0),
is given by u(x; t) = f (x −bt) for all t ≥ 0 and all x ∈ R
n
.
9.2. Method of Characteristics: More Examples
When using the method of characteristics, we exploit the fact that solutions of our
ﬁrstorder linear pdes must remain constant along certain curves in the xtplane.
In the case of the transport equation (c, 1) • ∇u(x, t) = 0, the characteristic curves
are straight lines parallel to the vector (c, 1). The lines have slope 1/c in the
xtplane (see Figure 9.2) and solutions u(x, t) must remain constant along these
242 method of characteri sti cs: more examples
t
x
1
c
(x−ct, 0)
(x,t)
Figure 9.2. Characteristic curves for the transport equation u
t
+ cu
x
= 0
are lines parallel to the vector (c, 1), shown in bold. Starting from any
point (x, t), following the characteristic curve backwards in time to t = 0
lets us express u(x, t) in terms of the initial condition.
lines. Now suppose t > 0 and let (x, t) be any point in the xtplane. Then to
determine u(x, t), we need only trace backwards in time along a characteristic
line until we reach t = 0. Indeed, since u(x, t) must remain constant as we follow
lines of slope 1/c, it must be the case that u(x, t) = u(x −ct, 0). If we are given
an initial condition u(x, 0) = f (x), this implies that u(x, t) = f (x −ct), which
we recognize as the solution of the Cauchy problem (i.e., initial value problem)
for the transport equation. We now use the method of characteristics to solve a
linear, homogeneous, ﬁrstorder pde with variable coefﬁcients.
A partial differential equation with a variable coefﬁcient. Consider the pde
u
t
+ xu
x
= 0, which is equivalent to ∇u(x, t) • (x, 1) = 0. The latter expression
says that the directional derivative of u vanishes in the direction of the vector
(x, 1). Notice that as x varies, so does the orientation of the vector (x, 1), and
sketching this vector ﬁeld (direction ﬁeld) helps us visualize the characteristic
curves. More exactly, the characteristic curves are the curves in the xtplane
that have (x, 1) as their tangent vectors. Since these vectors have slope 1/x, the
characteristic curves satisfy the ordinary differential equation
dt
dx
=
1
x
.
li near, fi rst order parti al di fferenti al equati ons 243
Assuming x ,= 0, we may use separation of variables to solve this ode:
_
1 dt =
_
1
x
dx,
which implies that t = ln x + C, where C is a constant. The family of character
istic curves in the xtplane is shown in Figure (9.3). Since a solution u(x, t) of
the pde must remain constant as we move along a characteristic curve, the value
of u(x, t) is completely determined by the value of the constant C = t −ln x if
x ,= 0. Therefore, any function of the form u(x, t) = g(t −ln x) is a solution
of the pde u
t
+ xu
x
= 0 whenever x ,= 0. To determine what happens when
x = 0, notice that the vector (0, 1) is always tangent to the characteristic curve
when x = 0. This implies that the vertical axis x = 0 in the xtplane is also a
characteristic curve which, in turn, implies that the value of u(0, t) can never
change as t varies.
We make several important observations concerning this example:
The characteristic curves completely ﬁll the xtplane without intersecting
each other.
Each characteristic curve intersects the horizontal axis t = 0. Therefore, if we
are provided with an initial condition u(x, 0) = f (x), then we can determine the
value of u(x, t) by following the characteristic curve through (x, t) until we reach
t = 0.
Every characteristic curve (excluding x = 0) is the graph of an invertible
function.
The latter two observations can be used to give a clean solution of the Cauchy
problem for this pde. Suppose that (x, t) is any point in the xtplane and, for
convenience, assume x > 0. The point (x, t) lies on some characteristic curve,
t = ln x + C where C is a constant. Following this curve to t = 0, we ﬁnd that
0 = ln x + C, which implies that x = e
−C
when t = 0. Therefore, since u(x, t)
remains constant as we follow a characteristic curve, it must be the case that
u(x, t) = u(e
−C
, 0). Finally, since C = t −ln x, we have shown that
u(x, t) = u
_
e
ln x−t
, 0
_
= u
_
xe
−t
, 0
_
= f (xe
−t
).
The case x < 0 is handled similarly. Finally, if x = 0 we have u(x, 0) = f (0),
implying that u must remain constant when x = 0 (which we already knew).
244 method of characteri sti cs: more examples
t
x
Figure 9.3. The family of characteristic curves for the equation u
t
+
xu
x
= 0 consists of logarithmic functions t = ln x + C and the vertical
axis x = 0.
We have shown that the solution of the initial value problem u
t
+ xu
x
= 0 with
u(x, 0) = f (x) is given by u(x, t) = f (xe
−t
).
A graphical interpretation of this solution can be given as follows. As t
increases from 0, the exponential function e
−t
decays from 1 to 0. This has
the effect of dilating the initial distribution of u, stretching it horizontally as
t increases. For example, when t = ln2, the solution is given by u(x, ln2) =
f (x/2). The graph of f (x/2) has the same qualitative appearance as the initial
condition f (x), but stretched horizontally outward (from x = 0) by a factor of 2.
Example 9.2.1. Solve the pde
2u
x
+3x
2
y
2
u
y
= 0 (−∞ < y < ∞)
with the auxiliary condition u(0, y) = e
−y
. Solution: Writing the pde in the form
∇u(x, y) • (2, 3x
2
y
2
) = 0, we see that the characteristic curves must satisfy the
ode
dy
dx
=
3
2
x
2
y
2
.
Separation of variables leads us to
_
1
y
2
dy =
3
2
_
x
2
dx,
li near, fi rst order parti al di fferenti al equati ons 245
and performing the integration reveals that
−
1
y
=
1
2
x
3
+ C,
where C is a constant. By algebra, the characteristic curves are given by
x =
3
¸
−
2
y
−2C.
Since u(x, y) must remain constant along characteristic curves, the value of
u(x, y) depends only upon the value of C = −
1
y
−
1
2
x
3
. Therefore, the general
solution of the pde is
u(x, y) = f
_
−
1
y
−
1
2
x
3
_
,
where f is an arbitrary differentiable function. Setting x = 0 allows us to use the
auxiliary condition u(0, y) = e
−y
:
f
_
−
1
y
_
= e
−y
.
Making the substitution z = −
1
y
provides the explicit formula
f (z) = e
1/z
.
Finally, combining this formula for f with the general solution of the pde yields
the overall solution
u(x, y) = exp
_
−
1
1
y
+
1
2
x
3
_
,
where we have written exp(z) instead of e
z
for notational convenience.
The preceding example merits a few general remarks summarizing the method
of characteristics. For a(x, t)u
t
+ b(x, t)u
x
= 0, a linear, homogeneous, ﬁrst
order pde, the associated characteristic equation is given by
dt
dx
=
a(x, t)
b(x, t)
or
dx
dt
=
b(x, t)
a(x, t)
.
Ideally, this ode can be solved by hand, although this is certainly not the case if
the coefﬁcient functions a(x, t) and b(x, t) are too complicated. The solutions of
246 method of characteri sti cs: more examples
the ode form the family of characteristic curves. Assuming that the characteristic
curves do not intersect one another (as in the above examples), the solution
u(x, t) is completely determined by which characteristic curve the pair (x, t) lies
on. In such cases, it is desirable that all characteristic curves exist for all t ≥ 0.
That way, determining the solution u(x, t) when t > 0 is a simple matter of
following the characteristic curve through (x, t) backwards in time until we reach
t = 0, thereby allowing us to invoke any initial conditions. Unfortunately, there
are many simplelooking ﬁrstorder pdes giving rise to characteristic curves that
are far from ideal.
Example 9.2.2. The ﬁrstorder nonlinear pde u
t
+ uu
x
= 0 is known as Burg
ers’ Equation. Equivalently, ∇u(x, t) • (u(x, t), 1) = 0, which leads us to the
characteristic equation
dx
dt
= u(x, t).
Without knowing the solutions u(x, t) of the pde, it would appear that solving
this ode is impossible. However, we actually can learn quite a bit about the
characteristic curves. Suppose that (x(t), t) is a parametrization of a characteristic
curve. To see that u remains constant along this curve, we ﬁrst use the chain rule
to calculate
d
dt
u(x(t), t) =
∂u
∂x
dx
dt
+
∂u
∂t
.
However, we know that
dx
dt
= u(x, t) according to the above characteristic ode.
Therefore,
d
dt
u(x(t), t) = u
t
+ uu
x
= 0,
which means that u(x(t), t) is constant. Since u must remain constant along each
characteristic curve, the ode
dx
dt
= u(x, t) reduces to
dx
dt
= constant. In other
words, we have shown that the characteristic curves are actually straight lines.
The slope of the characteristic curve through the point (x, t) is equal to the value
of u(x, t).
To explore the consequences of these observations, suppose that we impose
the initial condition
u(x, 0) = φ(x) =
_
¸
¸
¸
_
¸
¸
¸
_
1 if x ≤ −1
−x if −1 < x < 0
0 if 0 ≤ x.
(9.5)
li near, fi rst order parti al di fferenti al equati ons 247
−1 0 1
x
t
Figure 9.4. The family of characteristic curves for the equation u
t
+
uu
x
= 0 for the initial condition (9.5).
To sketch the characteristic curves in the xtplane, it helps to rewrite the charac
teristic ode as
dt
dx
=
1
u(x, t)
,
with
dt
dx
undeﬁned if u(x, t) = 0. According to our initial condition, the char
acteristic curves have slope
dt
dx
= 1 for x ≤ −1, slope −1/x for −1 < x < 0,
and undeﬁned slope for x ≥ 0. A sketch of the characteristic curves appears in
Figure 9.4. Notice that the lines intersect one another at t = 1, a feature of this
pde that has some rather interesting implications. The slopes of the characteristic
lines determine the velocity at which our initial distribution of u(x, t) will prop
agate. For the initial condition (9.5), the propagation velocity is 0 when x > 0
but is positive when x < 0. Three snapshots of the solution u(x, t) are shown in
Figure 9.5. The wave moves from left to right, and the regions where u is highest
have the fastest propagation speed. At time t = 1, the spatial proﬁle of u(x, t)
develops a discontinuous jump at x = 0. Examining how the characteristic lines
cross each other helps explain the formation of this shock wave. The shock forms
when the fast moving parts of the wave encounter the “stationary” region x > 0.
Describing what happens after the shock is formed is a subject more suitable for
an advanced pde course.
248 method of characteri sti cs: more examples
t = 0 t = 0.5 t = 1
−1 0 0 −1
x
−1 0
u(x,t) u(x,t) u(x,t)
x x
Figure 9.5. Solution of u
t
+ uu
x
= 0 with initial condition (9.5) at times
t = 0 (the initial condition), t = 1/2 and t = 1. A shock has formed at
t = 1.
Exercises
1. Solve the Cauchy problem
2u
t
+5u
x
= 0
u(x, 0) =
1
1 + x
2
and sketch u(x, t) versus x when t = 0, t = 5 and t = 10.
2. Use the substitution v = u
y
to ﬁnd the general solution of 2u
y
+ u
xy
= 0.
3. Solve the pde (e
x
+ e
−x
)u
t
+ 2u
x
= 0, and sketch a few of the characteristic
curves.
4. Let g be a function of one variable and let G denote an antiderivative of
g. Use the method of characteristics to show that the general solution of
g(x)u
t
+ u
x
= 0 is given by u(x, t) = f (t − G(x)), where f is an arbitrary
differentiable function.
5. Let g and h be a functions of one variable and let G and H denote antideriva
tives of g and h, respectively. Use the method of characteristics to show that the
general solution of g(x)u
t
+ h(t)u
x
= 0 is given by u(x, t) = f (H(t) −G(x)),
where f is an arbitrary differentiable function.
6. Solve the pde u
t
+2tx
2
u
x
= 0 with initial condition u(x, 0) = sin x.
7. Solve the initial value problem
(1 + t
2
)u
t
+ u
x
= 0 (−∞ < x < ∞),
u(x, 0) = e
−x
2
.
li near, fi rst order parti al di fferenti al equati ons 249
Then, sketch three different characteristic curves. Finally, plot u(x, t) versus x
for three different choices of t.
8. These questions concern the initial value problem for the inhomogeneous
transport equation on the whole real line:
u
t
+ cu
x
= g(x, t) (−∞ < x < ∞),
u(x, 0) = f (x).
It is possible to show that the solution of this initial value problem is
u(x, t) = f (x −ct) +
_
t
0
g (x + c(s −t), s) ds. (9.6)
Note that the inhomogeneity g(x, t) gives rise to an integral which was not
present when we solved the homogeneous transport equation.
(a) Using Equation (9.6), show that the solution of the initial value problem
u
t
+ u
x
= x cos t, (−∞ < x < ∞)
u(x, 0) =
1
1 + x
2
is given by
u(x, t) =
1
1 + (x −t)
2
+ x sin t + cos t − 1.
(b) Now, using direct substitution, check that u(x, t) really does satisfy the
initial value problem in Part (a).
CHAPTER 10
The Heat and Wave Equations on an Unbounded Domain
A
t ﬁrst glance, the heat equation u
t
− κu
xx
= 0 and the wave equation
u
tt
−c
2
u
xx
= 0 appear very similar. Since κ and c
2
are always assumed
to be positive constants, the only apparent distinction is between the u
t
in the
heat equation and the u
tt
in the wave equation. As we shall see, this makes
a profound difference in the behavior of the solutions of these two equations.
We begin this Chapter with a derivation of these two pdes from basic physical
principles. Then, we will solve both pdes on the unbounded onedimensional
spatial domain −∞ < x < ∞.
10.1. Derivation of the Heat and Wave Equations
The heat equation arises in the context of modeling diffusive processes. For
example, suppose heat is distributed within a long, thin wire that can be treated
as onedimensional. Further suppose that the wire is insulated so that heat is only
transferred within the wire (as opposed to radially outward). If T(x, t) denotes
the temperature at position x and time t, then T satisﬁes the heat equation.
The Heat Equation. For the purposes of deriving the heat equation, we actually
have in mind another phenomenon that is modeled by the same pde. Namely,
suppose that a dye (such as food coloring) diffuses in a motionless liquid that
is conﬁned to a long, thin “onedimensional” pipe. Letting u(x, t) denote the
concentration of dye at position x and time t, our goal is to derive an equation
that captures the dynamics of u. Generally, dye diffuses from regions of higher
concentration to regions of lower concentration, and the relevant physical law
that governs diffusion is
250
the heat and wave equati ons on an unbounded domai n 251
Fick’s Law of Diffusion. The rate of motion of dye is proportional to the con
centration gradient.
Fick’s Law will let us track how the total mass of dye in a given region of pipe,
say [x
0
, x
1
], changes over time. Mass is obtained by integrating concentration;
i.e., the mass of dye in the region [x
0
, x
1
] at time t is given by
M(t) =
_
x
1
x
0
u(x, t) dx.
Differentiating with respect to t yields
dM
dt
=
_
x
1
x
0
u
t
(x, t) dx.
Since
dM
dt
measures the rate of change of mass, it is equal to the difference
between the rate of ﬂow of dye into the interval [x
0
, x
1
] and the rate of ﬂow
of dye out of the interval [x
0
, x
1
]. By Fick’s Law, the rate of ﬂow of the dye is
proportional to the (spatial) concentration gradient u
x
. Letting κ > 0 denote
the proportionality constant, the rate of inward ﬂow at the left endpoint x
0
is
given by −κu
x
(x
0
, t). To explain why the negative sign is included, suppose
u
x
(x
0
, t) > 0. Then the concentration is higher for x > x
0
than for x < x
0
, which
implies that the ﬂow at x
0
would be from righttoleft (see Figure 10.1). By
contrast, if u
x
(x
0
, t) < 0, then the concentration is higher for x < x
0
than for
x > x
0
, thereby resulting in a ﬂow from lefttoright at x
0
. Similar reasoning
shows that the rate of inward ﬂow at the right endpoint x
1
is given by κu
x
(x
1
, t).
Therefore,
dM
dt
= κu
x
(x
1
, t) −κu
x
(x
0
, t),
and equating this to our original expression for
dM
dt
yields
_
x
1
x
0
u
t
(x, t) dx = κu
x
(x
1
, t) −κu
x
(x
0
, t).
Taking the derivative of both sides with respect to x
1
, we see that
u
t
(x
1
, t) = κu
xx
(x
1
, t).
252 deri vati on of the heat and wave equati ons
x
1
x
0
u
x
< 0
u
x
> 0
u(x,t)
x
Figure 10.1. Equation (10.1) describes diffusive processes in which
diffusion is proportional to the concentration gradient u
x
. In this ﬁgure,
u
x
> 0 at the left endpoint x
0
, and diffusion should cause a net outward
ﬂow (from right to left) at x
0
. Similarly, u
x
< 0 at the right endpoint x
1
.
The ﬂow should be from left to right at x
1
, i.e., from higher concentration
towards lower concentration. Note that the direction of motion has the
“opposite sign” of the gradient u
x
.
Since x
1
was arbitrary, we have derived the heat equation or diffusion equation
u
t
= κu
xx
. (10.1)
The Wave Equation. We now use physical principles to derive the wave equation
u
tt
−c
2
u
xx
= 0. Whereas the heat equation was derived from Fick’s Law, the
wave equation can be derived from Newton’s second law of motion, F = ma.
Suppose that we pluck a ﬂexible, elastic guitar string of uniform density. Our
goal is to completely describe the displacement u(x, t) from the equilibrium
position of the string (see Figure 10.2). We will make several assumptions con
cerning the dynamics of the vibrating string. First, we assume that all vibrations
are transverse (i.e., upanddown in Figure 10.2) as opposed to longitudinal (i.e.,
sidetoside in the ﬁgure). Second, we assume that the vibrations are “small” in
the sense that both u(x, t) and u
x
(x, t) have low magnitudes. Finally, we assume
that the string has constant density ρ and that air resistance is negligible.
Let T(x, t) denote the tension (force) vector at position x and time t. Tension
T(x, t) is directed tangent to the string, and therefore has the same direction as the
vector (1, u
x
(x, t)) (see Figure 10.2). We will examine the force and acceleration
on a region [x
0
, x
1
] of the guitar string. The idea is to write Newton’s Law
the heat and wave equati ons on an unbounded domai n 253
F = ma in terms of its transverse and longitudinal components where, in this
case, the force vector F is tension. Mathematically,
T(x, t) = T
long
(x, t) +T
trans
(x, t).
Since T(x, t) is parallel to the vector (1, u
x
(x, t)), we can derive relationships
between the lengths of transverse and longitudinal components of the vectors by
using a “similar triangles” argument (see Figure 10.3). Letting T denote the
length of the tension vector, we see that
T
long

T
=
1
_
1 + u
2
x
,
from which it follows that the magnitude of the longitudinal force is given by
T
long
 =
T
_
1 + u
2
x
.
Similarly,
T
trans

T
=
u
x
_
1 + u
2
x
,
from which it follows that the magnitude of the transverse force is given by
T
trans
 =
Tu
x
_
1 + u
2
x
. (10.2)
By assumption, the longitudinal acceleration of the region [x
0
, x
1
] is 0, and
therefore
T(x, t)
_
1 + u
2
x
¸
¸
¸
¸
¸
x
1
x
0
= 0. (10.3)
Since u(x, t) represents transverse displacement, the transverse acceleration is
u
tt
(x, t), the second derivative of u with respect to t. If we partition the interval
[x
0
, x
1
] into small segments of width ∆x, then the mass of each segment is
ρ∆x. Letting x
∗
j
denote the right endpoint of the j
th
segment, the transverse
acceleration of the j
th
segment is approximately u
tt
(x
∗
j
, t). Transverse force on
the j
th
segment is given by Newton’s law
F
trans
= ma
trans
≈ ρu
tt
(x
∗
j
, t)∆x.
254 deri vati on of the heat and wave equati ons
Summing over all segments yields a Riemann sum for the transverse force exerted
on the whole interval [x
0
, x
1
], namely
F
trans
≈
∑
j
ρu
tt
(x
∗
j
, t)∆x.
As ∆x → 0, this Riemann sum converges to an integral
F
trans
=
_
x
1
x
0
ρu
tt
(x, t) dx.
Since the transverse force is due to tension, this expression for transverse force
must match our earlier expression (10.2):
T(x, t)u
x
_
1 + u
2
x
¸
¸
¸
¸
¸
x
1
x
0
=
_
x
1
x
0
ρu
tt
(x, t) dx. (10.4)
Since we have assumed that the transverse vibrations are small, we know that
u
x
 ≈ 0. Thus, we may approximate
_
1 + u
2
x
≈ 1 with a high degree of accuracy.
This simpliﬁcation reduces the longitudinal equation (10.3) to
T(x, t)
√
1
¸
¸
¸
¸
x
1
x
0
= 0,
which means that T(x
1
, t) = T(x
0
, t). In other words, tension has approxi
mately constant magnitude along the entire vibrating string, and so T is actually
independent of x. It is also reasonable to assume that the tension is independent
of time t, and from now on we will let T denote the (constant) tension force.
With these simpliﬁcations, the equation (10.4) for transverse force reduces to
Tu
x
¸
¸
¸
¸
x
1
x
0
=
_
x
1
x
0
ρu
tt
(x, t) dx.
Differentiating with respect to x
1
and applying the Fundamental Theorem of
Calculus, we are left with Tu
xx
= ρu
tt
. Introducing the constant c =
_
T/ρ, we
have derived the wave equation
u
tt
−c
2
u
xx
= 0. (10.5)
the heat and wave equati ons on an unbounded domai n 255
u(x,t)
x
T(x,t)
Figure 10.2. The displacement u(x, t) of a guitar string (solid curve)
from its resting position (dashed line) at a ﬁxed time t. The tension
vector T(x, t) is tangent to the curve.
T
trans
 
u
x
1 + u
x
2
T
long
 
1
Figure 10.3. Longitudinal and transverse components of the tension
vector T. The hypotenuse of the larger triangle has length T, the
magnitude of the tension vector.
10.2. Cauchy Problem for the Wave Equation
In this section we will solve the wave equation (10.5) on the unbounded domain
−∞ < x < ∞. Before imposing initial conditions, we will derive the general
solution of the equation.
General solution. Starting from the wave equation
_
∂
2
∂t
2
−c
2
∂
2
∂x
2
_
u = 0,
the idea is to factor the operator as
_
∂
∂t
−c
∂
∂x
__
∂
∂t
+ c
∂
∂x
_
u = 0. (10.6)
256 cauchy problem for the wave equati on
Note that separately, the operators
_
∂
∂t
−c
∂
∂x
_
and
_
∂
∂t
+ c
∂
∂x
_
remind us of the transport equation. Since we know that the general solution of
the transport equation
_
∂
∂t
+ c
∂
∂x
_
u = 0
is given by u(x, t) = g(x −ct) where g is an arbitrary differentiable function, we
claim that the general solution of the wave equation (10.6) is given by
u(x, t) = f (x + ct) + g(x −ct), (10.7)
where f and g are arbitrary differentiable functions. To prove this claim, we write
the secondorder pde (10.6) as a system of two ﬁrstorder pdes by introducing
the new variable
w =
_
∂
∂t
+ c
∂
∂x
_
u = u
t
+ cu
x
.
The resulting system is
w
t
−cw
x
= 0 and u
t
+ cu
x
= w.
The solution of the homogeneous equation for w is w(x, t) = h(x + ct), where h
is an arbitrary differentiable function. Substituting this formula for w into the
other equation yields
u
t
+ cu
x
= h(x + ct).
How should we go about solving this inhomogeneous pde? As with odes, one
option is to produce one particular solution of the inhomogeneous equation
and add it to the general solution of the associated homogeneous pde. The
homogeneous problem u
t
+ cu
x
= 0 has general solution u(x, t) = g(x − ct),
where g is an arbitrary differentiable function. To produce a particular solution of
the inhomogeneous equation, it is natural to guess that if H is an antiderivative
for h, then H(x + ct) will satisfy the pde. This is almost the case—suppose that
u(x, t) = H(x + ct). Then
u
t
= cH
/
(x + ct) = ch(x + ct) and u
x
= H
/
(x + ct) = h(x + ct).
the heat and wave equati ons on an unbounded domai n 257
Therefore, u
t
+ cu
x
= 2ch(x + ct), and we see that our guess was merely off by a
factor of 2c. This is easily remedied by rescaling our guess; speciﬁcally, deﬁne
f (ξ) =
1
2c
H(ξ) =
1
2c
_
h(ξ) dξ,
a scalar multiple of the antiderivative of h. It is straightforward to check that
u(x, t) = f (x + ct) actually is a particular solution of u
t
+ cu
x
= h(x + ct).
Adding this particular solution to the general solution of the homogeneous
equation u
t
+ cu
x
= 0 establishes that (10.7) really is the general solution of the
wave equation.
Notice that the general solution of the wave equation contains two arbitrary
functions, which is no coincidence since the wave equation is secondorder in the
variable t. If c > 0, then the function f (x + ct) corresponds to a wave moving
righttoleft with speed c, and the function g(x − ct) corresponds to a wave
moving lefttoright with speed c. Whereas the transport equation u
t
+ cu
x
= 0
gives rise to one family of characteristic curves (x −ct = constant), the wave
equation gives rise to two families of characteristic curves: x ± ct = constant. In
other words, the wave equation transmits waves in both directions, whereas the
transport equation can only transmit a wave in one direction.
The Cauchy Problem. Recall that a Cauchy problem for a pde is essentially an
initial value problem on an unbounded domain (i.e., boundary conditions are not
needed). Since the wave equation models the displacement of a vibrating guitar
string from its resting position, intuitively we expect that two initial conditions
will be needed: initial displacement and initial velocity. Indeed, the Cauchy
problem for the wave equation is given by
u
tt
−c
2
u
xx
= 0 (−∞ < x < ∞),
u(x, 0) = φ(x) (−∞ < x < ∞),
u
t
(x, 0) = ψ(x) (−∞ < x < ∞),
(10.8)
where φ(x) and ψ(x) represent the initial displacement and velocity of the string,
respectively. Since we know the general solution of the wave equation is given by
u(x, t) = f (x + ct) + g(x −ct),
258 cauchy problem for the wave equati on
we must incorporate the initial conditions to solve for f and g in terms of φ and
ψ. Setting t = 0 yields
u(x, 0) = φ(x) = f (x) + g(x).
To use the initial condition for velocity, differentiate with respect to t to obtain
u
t
(x, t) = c f
/
(x + ct) −cg
/
(x −ct),
and then set t = 0:
u
t
(x, 0) = ψ(x) = c f
/
(x) −cg
/
(x).
Integrating ψ = c f
/
−cg
/
over the interval [0, x],
1
c
_
x
0
ψ(s) ds =
_
x
0
f
/
(s) −g
/
(s) ds = f (x) −g(x) − f (0) + g(0).
Combined with our earlier equation φ = f + g, we obtain the system
f (x) + g(x) = φ(x) (10.9)
f (x) −g(x) = f (0) −g(0) +
1
c
_
x
0
ψ(s) ds. (10.10)
Adding (10.9) and (10.10) and solving for f (x), we have
f (x) =
1
2
φ(x) +
1
2c
_
x
0
ψ(s) ds +
f (0) −g(0)
2
, (10.11)
and subtracting (10.10) from (10.9) yields
g(x) =
1
2
φ(x) −
1
2c
_
x
0
ψ(s) ds −
f (0) −g(0)
2
. (10.12)
Substituting the expressions (10.11) and (10.12) into the general solution (10.7)
of the wave equation,
u(x, t) = f (x + ct) + g(x −ct) =
1
2
φ(x + ct) +
1
2c
_
x+ct
0
ψ(s) ds
+
1
2
φ(x −ct) −
1
2c
_
x−ct
0
ψ(s) ds.
(10.13)
the heat and wave equati ons on an unbounded domai n 259
(Notice that all of the stray constants in (10.11) and (10.12) have now dropped
out.) Combining the two integrals in (10.13), we have shown that
u(x, t) =
1
2
[φ(x + ct) + φ(x −ct)] +
1
2c
_
x+ct
x−ct
ψ(s) ds (10.14)
is the solution of the Cauchy problem for the wave equation. We will refer to it
as D’Alembert’s formula in honor of the mathematician who ﬁrst discovered it.
By the very construction of D’Alembert’s formula (10.14), it is evident that it
represents the unique solution of the Cauchy problem for the wave equation, and
that small changes in φ and ψ do not cause major changes in the behavior of
solutions. In this sense, the Cauchy problem for the wave equation is wellposed—
there is precisely one solution and it is not hypersensitive to small changes in
the initial conditions.
Inspection of D’Alembert’s formula (10.14) provides some insight regarding
how solutions of the wave equation should behave. If the initial velocity is zero,
then the integral term vanishes and we are left with u(x, t) =
1
2
[φ(x +ct) +φ(x −
ct)]. Effectively, this says that our initial displacement φ will be “split” into two
waves traveling in opposite directions with speed c and with half the amplitude
of the initial wave proﬁle. The next example reinforces this observation.
Example 10.2.1. Solve the Cauchy problem
u
tt
−u
xx
= 0 (−∞ < x < ∞),
u(x, 0) = φ(x) (−∞ < x < ∞),
u
t
(x, 0) = 0 (−∞ < x < ∞),
where the initial displacement φ(x) is given by φ(x) = 1 if −1 ≤ x ≤ 1 and
φ(x) = 0 otherwise.
Solution: Although it may seem odd that the initial displacement is discon
tinuous, nothing prevents us from quoting D’Alembert’s formula. In this case,
the wave speed is c = 1 and the initial velocity is ψ = 0. According to (10.14), the
solution of this initial value problem is u(x, t) =
1
2
[φ(x + t) + φ(x −t)], where φ
is the given initial displacement. By deﬁnition of φ, we have
φ(x + t) =
_
_
_
1 if −1 ≤ x + t ≤ 1
0 otherwise
260 cauchy problem for the wave equati on
and
φ(x −t) =
_
_
_
1 if −1 ≤ x −t ≤ 1
0 otherwise.
It is instructive to sketch the domains −1 ≤ x + t ≤ 1 and −1 ≤ x −t ≤ 1 in the
xt plane (see Figure 10.4). The Figure shows the regions in which φ(x + t) = 1,
φ(x −t) = 1, both are 1, or both are 0.
Example 10.2.2. Solve the Cauchy problem
u
tt
−9u
xx
= 0 (−∞ < x < ∞),
u(x, 0) = e
−x
2
(−∞ < x < ∞),
u
t
(x, 0) = xe
−x
2
(−∞ < x < ∞).
Solution: In this case, the speed is c = 3. Quoting D’Alembert’s formula, the
solution is given by
u(x, t) =
1
2
_
e
−(x+3t)
2
+e
−(x−3t)
2
_
+
1
6
_
x+3t
x−3t
se
−s
2
ds.
A substitution v = −s
2
, −
1
2
dv = sds facilitates the integration:
u(x, t) =
1
2
_
e
−(x+3t)
2
+e
−(x−3t)
2
_
+
1
6
_
−(x+3t)
2
−(x−3t)
2
−
1
2
e
v
dv.
Evaluating the integral, the overall solution is given by
u(x, t) =
1
2
_
e
−(x+3t)
2
+e
−(x−3t)
2
_
−
1
12
_
e
−(x+3t)
2
−e
−(x−3t)
2
_
=
5
12
e
−(x+3t)
2
+
7
12
e
−(x−3t)
2
.
Notice that in this example, the nonzero initial velocity breaks the “symmetry”
of the two waves that propagate outward. In some sense, we see that
7
12
of our
initial displacement proﬁle travels to the right with speed 3, while only
5
12
travels
to the left with speed 3. This may seem a bit surprising considering that both the
initial displacement and velocity have even and odd symmetries, respectively.
the heat and wave equati ons on an unbounded domai n 261
−1
1
x
−0.5 −1.5 1.5 0.5
1
0.5 0.5
−1.5 1.5
1
0.5
x
x
x
0.5 0.5
−3 −1 1 3
u(x,0) u(x,0.5)
u(x,2) u(x,1)
φ(x+t)=1 φ(x−t)=1
x
t
−1
0 1
Both one
Both zero
x+t=−1 x+t=1 x−t=−1 x−t=1
Figure 10.4. Wave propagation with a discontinuous initial displace
ment φ(x) (see example in text). Upper ﬁgure: Regions in which either
φ(x + t) = 1, φ(x − t) = 1, both are 1, or both are 0. The dashed
horizontal lines, included for reference, correspond to t = 0.5, 1 and 2.
Lower ﬁgure: Four frames of an “animation” of the solution. The initial
condition u(x, 0) is discontinuous, with jumps occurring at x = ±1. At
time t = 1/2, we see that the initial displacement is beginning to split
into two waves of half amplitude propagating outward. At t = 1, the
two waves have almost separated from each other, and by t = 2 we
have two completely separate waves. The two waves are identical to the
initial displacement but have half the amplitude.
262 cauchy problem for the wave equati on
u(x,0)
x
Figure 10.5. Initial velocity u
t
(x, 0) = −cos x (indicated by arrows) acts
opposite the initial displacement u(x, 0) = cos x.
Example 10.2.3. Solve the Cauchy problem
u
tt
−c
2
u
xx
= 0 (−∞ < x < ∞),
u(x, 0) = cos x (−∞ < x < ∞),
u
t
(x, 0) = −cos x (−∞ < x < ∞).
Solution: Before quoting D’Alembert’s formula, examine the initial conditions
carefully. The initial velocity acts completely “opposite” the initial displacement
as illustrated in Figure 10.5. Intuitively, this would seem to suggest that the
solution of the Cauchy problem will be a standing wave: the wave will oscillate
up and down, with motion perpendicular to the xaxis. Moreover, the string
should remain stationary at all the points where u(x, 0) = 0. By D’Alembert’s
formula (10.14), the solution of the Cauchy problem is
u(x, t) =
1
2
[cos(x + ct) +cos(x −ct)] −
1
2c
_
x+ct
x−ct
cos(s) ds
=
1
2
[cos(x + ct) +cos(x −ct)] −
1
2c
[sin(x + ct) −sin(x −ct)] .
(10.15)
Although this is the correct solution of our Cauchy problem, visualizing how
the solution behaves is not easy unless we write it in a simpler form. The
doubleangle identities from trigonometry
sin(α + β) = sin α cos β + cos α sin β
cos(α + β) = cos α cos β − sin α sin β
the heat and wave equati ons on an unbounded domai n 263
combined with the symmetry properties sin(−α) = −sin α and cos(−α) =
cos(α) will greatly simplify (10.15). The ﬁrst term in (10.15) is
1
2
[cos(x + ct) +cos(x −ct)]
=
1
2
[cos(x) cos(ct) −sin(x) sin(ct) +cos(x) cos(−ct) −sin(x) sin(−ct)]
=
1
2
[cos(x) cos(ct) −sin(x) sin(ct) +cos(x) cos(ct) +sin(x) sin(ct)]
= cos(x) cos(ct).
In the same way, we can simplify the other term in (10.15):
−
1
2c
[sin(x + ct) −sin(x −ct)] = −
1
c
cos(x) sin(ct).
Combining our two simpliﬁcations, our overall solution is now written as
u(x, t) = cos(x) cos(ct) −
1
c
cos(x) sin(ct) = cos(x)
_
cos(ct) −
1
c
sin(ct)
_
.
With u(x, t) expressed in this form, solutions are much easier to visualize. Notice
that the cos(x) factor corresponds to the initial displacement and depends only on
x. The timedependent factor cos(ct) −
1
c
sin(ct) experiences periodic oscillations
as t varies and, in doing so, modiﬁes the amplitude of the displacement. If we
use a computer to create an animation of u versus x as t increases, we would see
a standing wave. At points where cos(x) = 0, the string would appear to remain
stationary, whereas at all other points the string would continually oscillate up
and down.
Domains of Dependence and Inﬂuence. D’Alembert’s solution of the wave
equation u
tt
− c
2
u
xx
= 0 shows that the initial displacement proﬁle φ(x) ef
fectively splits into two waves traveling in opposite directions with speed c.
Certainly the initial velocity ψ(x) may inﬂuence the propagating waves, but
no part of the waves can travel with speed faster than c. There are two important
consequences of this remark. First, suppose that we select a point x
0
from within
our spatial domain −∞ < x < ∞. The characteristic curves through the point
(x
0
, 0) are the lines x = x
0
± ct, as illustrated in the upper panel of Figure 10.6.
The initial conditions φ(x
0
), ψ(x
0
) at the point x
0
cannot inﬂuence the behavior
264 cauchy problem for the wave equati on
x
o
x = x
o
− ct x = x
o
+ ct
x
t
x
t
domain of
influence
(x,t)
domain of
dependence
(x−ct,0) (x+ct,0)
Figure 10.6. Upper panel: The domain of inﬂuence for x
0
. The initial
condition at x
0
cannot affect how u(x, t) behaves outside the shaded
region because waves cannot travel faster than speed c. Lower panel:
The domain of dependence for u(x, t). If t > 0, the value of u(x, t)
can only depend upon the initial conditions within the closed interval
[x −ct, x + ct].
of the solution outside the shaded region in the xtplane, because the waves
cannot travel faster than speed c. For that reason, the shaded region is called the
domain of inﬂuence of x
0
. Likewise, if we select a point (x, t), it is natural to ask
for the range of xvalues for which the initial conditions could have any impact
on the value of u(x, t). The lower panel of Figure 10.6 illustrates the domain of
dependence for u(x, t)—i.e., the region in the xtplane which could inﬂuence the
value of u(x, t). In particular, because waves can travel no faster than speed c,
the value of u(x, t) can only depend upon the behavior of the initial conditions
within the interval [x −ct, x + ct] as shown in the Figure.
the heat and wave equati ons on an unbounded domai n 265
10.3. Cauchy Problem for the Heat Equation
We now turn our attention to the Cauchy problem for the heat equation:
u
t
−κu
xx
= 0 (−∞ < x < ∞) (10.16)
u(x, 0) = φ(x) (−∞ < x < ∞), (10.17)
where κ > 0 is a constant. In contrast with the Cauchy problem for the wave
equation, we have speciﬁed only one initial condition. Recalling the physical
intuition that guided our derivation of the heat equation, this is not surprising.
Indeed, we need only know the initial heat (or concentration) proﬁle of u versus
x, and then Fick’s Law will govern how the distribution of heat (or concentration)
will evolve over time.
Perhaps the most elegant way to attack the Cauchy problem for the heat
equation is by means of Fourier transforms. However, we do not presume that
the reader is acquainted with certain techniques from complex analysis that are
useful in understanding how the transform procedure works. Instead, we will
follow the text of Strauss [10], exploiting certain invariance properties of the heat
equation in order to construct the solution of (10.16) and (10.17). The idea is to
solve that system for a very special choice of the initial condition φ(x). Then, we
will use the solution of that special problem to construct the general solution
of the Cauchy problem. There are ﬁve properties of the heat equation and its
solutions that we will invoke:
Property 1: Translation invariance. If u(x, t) is a solution of (10.16), then for
any ﬁxed number x
0
, the function u(x −x
0
, t) is also a solution. Indeed, direct
differentiation yields u
t
(x −x
0
, t) = κu
xx
(x −x
0
, t).
Property 2: Derivatives of solutions. If u(x, t) satisﬁes the heat equation (10.16)
then the partial derivatives of u also satisfy the heat equation. For example, to
see that u
t
is also a solution, let v = u
t
. Then
v
t
= (u
t
)
t
= (κu
xx
)
t
= (κu
t
)
xx
= κv
xx
.
Thus, v = u
t
is also a solution.
266 cauchy problem for the heat equati on
Property 3: Linearity. Since the heat equation is linear, any ﬁnite linear combi
nation of solutions of the heat equation (10.16) is also a solution.
Property 4: Integrals and Convolutions. Suppose S(x, t) is a solution of the
heat equation (10.16). Then by translation invariance (Property 1), so is S(x −y, t)
for any ﬁxed choice of y. If g is any other function, we deﬁne the convolution of S
with g as
(S g)(x, t) =
_
∞
−∞
S(x −y, t)g(y) dy,
provided that this improper integral converges. We claim that, regardless of g,
the convolution (S g)(x, t) is also a solution of the heat equation. To check this,
we need to show that (S g)
t
−κ(S g)
xx
= 0. Taking the derivatives of the
convolution,
(S g)
t
−κ(S g)
xx
=
∂
∂t
_
∞
−∞
S(x −y, t)g(y) dy − κ
∂
2
∂x
2
_
∞
−∞
S(x −y, t)g(y) dy
=
_
∞
−∞
S
t
(x −y, t)g(y) dy −
_
∞
−∞
κS
xx
(x −y, t)g(y) dy
=
_
∞
−∞
[S
t
(x −y, t) −κS
xx
(x −y, t)]g(y) dy = 0,
as claimed.
Remark. In the above chain of calculations, we are only justiﬁed in moving the
derivatives under the integral sign provided that all of the functions involved
are reasonably wellbehaved. Certainly, all of the integrals need to converge.
For precise conditions under which interchanging derivatives and integrals is
justiﬁed, consult a textbook on mathematical analysis.
Property 5: Dilation. Suppose a > 0 is a constant. If u(x, t) is a solution of the
heat equation, then the dilated function v(x, t) = u(
√
a x, at) is also a solution. To
prove this, we calculate
∂
∂t
v(x, t) =
∂
∂t
u(
√
a x, at) = a u
t
(
√
a x, at)
∂
∂x
v(x, t) =
∂
∂x
u(
√
a x, at) =
√
a u
x
(
√
a x, at)
the heat and wave equati ons on an unbounded domai n 267
and
∂
2
∂x
2
v(x, t) = a u
xx
(
√
a x, at).
It follows that
v
t
(x, t) −κv
xx
(x, t) = a u
t
(
√
a x, at) −κa u
xx
(
√
a x, at)
= a
_
u
t
(
√
a x, at) −κ u
xx
(
√
a x, at)
¸
= 0,
which establishes that v is also a solution of the heat equation.
We now use these ﬁve Properties of the heat equation to build the general
solution of the Cauchy problem (10.16)–(10.17). Following Strauss [10], we ﬁrst
solve the heat equation with a special initial condition:
Q
t
−κQ
xx
= 0 (−∞ < x < ∞), (10.18)
Q(x, 0) = φ(x) =
_
_
_
1 if x > 0
0 if x ≤ 0
(−∞ < x < ∞). (10.19)
This seemingly strange initial condition φ(x) is very convenient because it is
dilation invariant: for any constant a > 0, the graph of φ(ax) is identical to that
of φ(x). Moreover, we also know (Property 5 above) that any solution Q(x, t)
of the heat equation is unaffected by the dilation x →
√
ax and t → at. These
observations have a nice implication: notice that the quantity x/
√
t is unaffected
by these dilations because
x
√
t
→
√
ax
√
at
=
x
√
t
.
Consequently, solutions Q(x, t) of (10.18)–(10.19) depend only on the quantity
x/
√
t. Therefore, we will seek solutions of the form
Q(x, t) = g
_
x
√
4κt
_
, (10.20)
where the constant 4κ included for later convenience. (The following computa
tions could be carried out without including the 4κ, but some equations would
look a bit messy.)
268 cauchy problem for the heat equati on
Now that we are seeking special solutions of the form (10.20) and g is a
function of one variable p = x/
√
4κt, we can reduce the heat equation to an
ordinary differential equation. Substituting Q(x, t) = g(p) into (10.18), the chain
rule yields
Q
t
=
dg
dp
∂p
∂t
=
∂p
∂t
g
/
(p).
By calculus and algebra,
∂p
∂t
=
∂
∂t
x
√
4κt
= −
1
2t
x
√
4κt
= −
p
2t
and, combining this with the previous equation, we have
Q
t
= −
1
2t
pg
/
(p).
Similarly, taking derivatives with respect to x reveals that
Q
x
=
dg
dp
∂p
∂x
=
1
√
4κt
g
/
(p)
and
Q
xx
=
dQ
x
dp
∂p
∂x
=
_
1
√
4κt
g
//
(p)
_
1
√
4κt
=
1
4κt
g
//
(p).
Combining these new expressions for Q
t
and Q
xx
with the heat equation (10.18),
we ﬁnd that
−
1
2t
pg
/
(p) −
κ
4κt
g
//
(p) = 0
for all t > 0. By algebra, this is equivalent to the secondorder variablecoefﬁcient
ode
g
//
(p) + 2p g
/
(p) = 0.
For emphasis, notice that we have effectively reduced the heat equation to a
secondorder ode by seeking solutions of a special type (10.20).
To solve the ode g
//
(p) +2pg
/
(p) = 0, we ﬁrst use algebra to write the equation
as
g
//
(p)
g
/
(p)
= −2p.
the heat and wave equati ons on an unbounded domai n 269
The sneaky observation that
g
//
(p)
g
/
(p)
=
d
dp
ln[g
/
(p)]
makes it easier to solve this ode. Indeed, integrating both sides of
d
dp
ln[g
/
(p)] = −2p
with respect to p reveals that
ln[g
/
(p)] = −p
2
+ C,
where C is a constant of integration. Exponentiating both sides, we have
g
/
(p) = e
−p
2
+C
= e
C
e
−p
2
= C
1
e
−p
2
,
where C
1
= e
C
is a constant. Integrating both sides,
g(p) = C
2
+
_
p
0
C
1
e
−r
2
dr,
where C
2
= g(0) is another constant (see Equation (2.16)). Thus far, we have
shown that the solution of the special problem (10.18)–(10.19) has the form
Q(x, t) = C
1
_
x/
√
4κt
0
e
−r
2
dr + C
2
. (10.21)
Note that this formula only makes sense for t > 0. As soon as we ﬁgure out
the values of the constants C
1
and C
2
, we will have solved the initial value
problem (10.18)–(10.19). Determining the values of these constants will require a
technical lemma.
Lemma 10.3.1.
_
∞
−∞
e
−x
2
dx =
√
π and
_
∞
0
e
−x
2
dx =
_
0
−∞
e
−x
2
dx =
√
π
2
.
270 cauchy problem for the heat equati on
Proof. Let I denote the value of the integral over the domain (−∞, ∞). Then
I
2
=
_
_
∞
−∞
e
−x
2
dx
__
_
∞
−∞
e
−y
2
dy
_
=
_
∞
−∞
_
∞
−∞
e
−(x
2
+y
2
)
dx dy.
Converting to polar coordinates,
I
2
=
_
2π
0
_
∞
0
re
−r
2
dr dθ.
The substitution ψ = r
2
simpliﬁes the interior integral:
I
2
=
_
2π
0
_
∞
0
1
2
e
−ψ
dψ dθ =
_
2π
0
1
2
dθ = π.
Therefore, I =
√
π, as claimed. The second statement in the Lemma follows from
the fact that the integrand has even symmetry.
With Lemma 10.3.1 in mind, we return to formula (10.21). If x > 0, then taking
the limit t → 0
+
yields
1 = lim
t→0
+
Q(x, t) = C
1
_
∞
0
e
−r
2
dr + C
2
.
The leftmost equality is a consequence of our initial condition and the fact that we
are temporarily assuming x > 0. Using Lemma 10.3.1, this statement simpliﬁes
to
1 =
√
π
2
C
1
+ C
2
.
Now suppose that x < 0, the region in which the initial condition is identically 0.
Taking the limit t → 0
+
as before, we have
0 = lim
t→0
+
Q(x, t) = C
1
_
−∞
0
e
−r
2
dr + C
2
= −C
1
_
0
−∞
e
−r
2
dr + C
2
.
Recognizing the integral from the Lemma, this equation reduces to
0 = −
√
π
2
C
1
+ C
2
.
the heat and wave equati ons on an unbounded domai n 271
Now that we have a system of two equations for C
1
and C
2
, namely
1 =
√
π
2
C
1
+ C
2
and 0 = −
√
π
2
C
1
+ C
2
,
routine algebra tells us that C
1
= 1/
√
π and C
2
= 1/2. Finally, we have shown
that the solution of the special Cauchy problem (10.18)–(10.19) is given by
Q(x, t) =
1
2
+
1
√
π
_
x/
√
4κt
0
e
−r
2
dr. (10.22)
Unfortunately, this integral cannot be simpliﬁed, because the function e
−r
2
does
not have an antiderivative that is expressible in terms of elementary functions.
Now that we have solved the Cauchy problem for the heat equation with a
special choice of initial conditions (10.19), we can combine our ﬁndings with the
ﬁve Properties to solve the general Cauchy problem (10.16)–(10.17).
All of our above work shows that the function Q(x, t) given by (10.22) is a
solution of the heat equation u
t
−κu
xx
= 0. By Property 2 (see above) of solutions
of the heat equation, we know that the derivative
S(x, t) =
∂Q
∂x
=
1
√
4πκt
e
−x
2
/4κt
is also a solution of the heat equation. By Property 1 (translation invariance),
it follows that S(x −y, t) is also a solution for each ﬁxed y. By Property 4, the
convolution of S(x, t) with the initial condition φ(x) is also a solution. That is,
(S φ)(x, t) =
_
∞
−∞
S(x −y, t)φ(y) dy
is also a solution of the heat equation. In fact, we shall soon see that this is the
unique solution of the Cauchy problem for the heat equation. For emphasis, the
solution of (10.16)–(10.17) is given by
u(x, t) = (S φ)(x, t) =
1
√
4πκt
_
∞
−∞
e
−(x−y)
2
/4κt
φ(y) dy. (10.23)
Notice that (10.23) only represents the solution for t > 0. Checking that this
formula obeys the initial condition u(x, 0) = φ(x) requires us to prove that
lim
t→0
+ u(x, t) = φ(x), a rather technical and tedious calculation.
272 cauchy problem for the heat equati on
Solving the Cauchy problem for the heat equation required substantial effort
and, unfortunately, the form of the solution (10.23) is not as “illuminating” as
what we might have hoped for. There are very few choices of initial conditions
φ for which it is possible to explicitly evaluate the integral in (10.23) by hand.
Certainly performing the integration with respect to the dummy variable y
should return a function of both x and t, but understanding the behavior of
solutions requires some thought (see below).
We make several additional remarks about the solution (10.23) before inter
preting that formula. First, observe that (10.23) makes no sense if t ≤ 0. (Keep
in mind that the diffusion constant κ was assumed to be positive.) Next, we
mention that S(x, t) has a special name:
Deﬁnition 10.3.2. The function
S(x, t) =
1
√
4πκt
e
−x
2
/4κt
(t > 0) (10.24)
is called the (onedimensional) heat kernel.
The heat kernel was deﬁned as S(x, t) = Q
x
(x, t), where Q was the solution
of the Cauchy problem with a special initial condition: a piecewise constant
function with a single jump discontinuity at x = 0. Then, we claimed that the
solution of the general Cauchy problem for the heat equation is obtained by
taking the convolution of the heat kernel S(x, t) with the initial condition φ(x).
Understanding why this mysterious process successfully solved the Cauchy
problem is a topic that is more suitable for an advanced (more theoretical) course
in pdes.
Now let us give a qualitative description of the behavior of the solution (10.23)
by analyzing the effects of the two factors that appear in the heat kernel (10.24).
For t > 0, the graph of the Gaussian function
e
−x
2
/4κt
is a “bell curve”. As t increases, the graph is dilated (stretched outward). The
other factor in the heat kernel, namely
1
√
4πκt
the heat and wave equati ons on an unbounded domai n 273
modulates the amplitude of the Gaussian curves. The amplitude blows up to ∞
as t → 0
+
and approaches 0 as t →∞. Combining the effects of the two factors
in the heat kernel, we see that S(x, t) is “tall and thin” for small positive t and
“short and broad” for large positive t. This is illustrated graphically in Figure 10.7.
Note that if y is a ﬁxed number, then the graph of S(x −y, t) is nothing more
than a horizontally shifted version of what appears in Figure 10.7 (the peaks
would be at x = y instead of at x = 0). As a ﬁnal remark about the heat kernel,
we mention that
_
∞
−∞
S(x, t) dx = 1
for all t > 0 (see Exercises), from which it follows that for any ﬁxed y,
_
∞
−∞
S(x −y, t) dx = 1
as well.
With our new understanding of the behavior of the heat kernel S(x, t), we
are in a position to interpret the solution of the Cauchy problem given by
formula (10.23). In the convolution
u(x, t) = (S φ)(x, t) =
_
∞
−∞
S(x −y, t)φ(y) dy,
the S(x −y, t) factor essentially gives a “weighted average” of the function φ,
exaggerating the points near y = x. As t increases, S(x −y, t) becomes shorter
and broader, but still has its peak at y = x. The convolution (S φ)(x, t) has the
effect of “diffusing” φ outward at each point in the spatial domain.
Example 10.3.3. The only initial condition for which we have explicitly calculated
the solution of the Cauchy problem for the heat equation is given by (10.19).
Using the Heaviside (unit step) function as the initial condition, we found that
the solution is given by
Q(x, t) =
1
2
+
1
√
π
_
x/
√
4κt
0
e
−r
2
dr.
Although the integrand does not have an elementary antiderivative, we could
use a computer to estimate the values of this integral for various choices of x and
t. More systematically, we could pick a speciﬁc t > 0 and then have a computer
approximate the integral for various choices of x. Results of such a procedure are
274 cauchy problem for the heat equati on
x
S
(
x
,
t
)
t = 0.1
t = 1.0
t = 5.0
Figure 10.7. Plot of the heat kernel S(x, t) versus x at three speciﬁc
times t = 0.1, t = 1.0, and t = 5.0. The diffusion coefﬁcient was chosen
to be κ = 1. Notice that as t increases, the graph of S(x, t) versus x
becomes shorter and broader.
shown in Figure 10.8, which shows plots of the initial condition Q(x, 0) as well
as the functions Q(x, 1) and Q(x, 2) versus x. Notice that jump discontinuity that
appeared in the initial condition is smoothed out. Moreover, the spatial proﬁle of
Q(x, t) is stretched/diffused outward as t increases, as described in the paragraph
preceding this example. In fact, despite the fact that the initial condition was
discontinuous, the function Q(x, t) is smooth (differentiable inﬁnitely many times)
for all t > 0. No matter how “ugly” the initial condition is, solutions of the heat
equation become instantly smooth. This is one of that many features of the heat
equation that distinguishes it from the wave equation.
Propagation speed. Another feature (or liability) of the diffusion/heat equation
is that “information” propagates with inﬁnite speed. To illustrate what we mean
by this, consider the Cauchy problem
u
t
−κu
xx
= 0
u(x, 0) = φ(x) =
_
_
_
1 if −1 ≤ x ≤ 1
0 otherwise
the heat and wave equati ons on an unbounded domai n 275
1
0
0 x
Q
(
x
,
t
)
t = 2
t = 1
t = 0
Figure 10.8. Plot of Q(x, t) versus x at three speciﬁc times t = 0, t = 1,
and t = 2.
on the usual domain −∞ < x < ∞ and t ≥ 0. Quoting formula (10.23), the
solution is given by
u(x, t) =
1
√
4πκt
_
∞
−∞
e
−(x−y)
2
/4κt
φ(y) dy =
1
√
4πκt
_
1
−1
e
−(x−y)
2
/4κt
dy.
Inspecting this integral, it is evident that for all real x, the value of u(x, t) is
positive when t > 0 because the exponential function in the integrand is always
positive. Of course, if x is huge, then the value of u(x, t) is barely positive. The
important thing to notice here is that although the initial condition is nonzero
only in the region x ∈ [−1, 1], this initial “disturbance” has an immediate (albeit
small) effect on the entire real line. The inﬁnite propagation speed associated
with the heat equation is not physically realistic, but the effect is small enough
that this pde is still a useful model of diffusive processes.
The preceding examples highlight some of the important distinctions between
solutions of the wave and heat equations. The Cauchy problem for the heat
equation is not wellposed for t ≤ 0, whereas D’Alembert’s solution of the
wave equation makes sense for all real t. If we use discontinuous initial data
for the wave equation, the discontinuities are preserved as t increases. By
contrast, discontinuities are instantly smoothed by the heat equation—solutions
are inﬁnitely differentiable for all t > 0. Finally, the issue of ﬁnite versus inﬁnite
propagation speed provides yet another fundamental distinction between waves
and diffusions.
276 well posedness and the heat equati on
10.4. WellPosedness and the Heat Equation
In the previous section, we found a solution (10.23) of the Cauchy problem for
the heat equation. In order for the Cauchy problem to be wellposed, existence of
a solution is not enough. We also need to know that (i) the solution is unique and
(ii) small changes in the initial condition φ(x) should not cause major changes
in how the solution behaves. We will use two different methods to prove that
initial/boundary value problems for the heat equation are wellposed. The ﬁrst
method, known as the energy method, is more versatile because it can be used
to prove wellposedness of the Cauchy, Dirichlet, and Neumann problems for
the heat equation as well as other pdes. The second method exploits a special
property of the heat equation known as the Maximum Principle, which offers
additional insight into the heat equation’s dynamics.
Energy Method. Consider the Cauchy problem
u
t
−κu
xx
= 0
u(x, 0) = φ(x)
on the usual domain −∞ < x < ∞ and t ≥ 0. In an above example, we argued
that the solution of this problem is smooth (inﬁnitely differentiable) for t > 0.
We now prove that there is a unique smooth solution provided that φ(x) → 0 as
x → ±∞. The assumption that φ vanishes as we approach ±∞ is actually not
very restrictive at all.
To prove uniqueness, suppose that u(x, t) and v(x, t) are solutions of the above
Cauchy problem. We must show that u = v or, equivalently, that w = u −v is
zero. Since both u
t
= κu
xx
and v
t
= κv
xx
, by linearity it follows that
w
t
= (u −v)
t
= u
t
−v
t
= κu
xx
−κv
xx
= κ(u −v)
xx
= κw
xx
.
Therefore, w is also a solution of the heat equation. As for the initial conditions,
since both u(x, 0) = φ(x) and v(x, 0) = φ(x), we have w(x, 0) = u(x, 0) −
v(x, 0) = 0. In summary, w satisﬁes the Cauchy problem
w
t
= κw
xx
and w(x, 0) = 0.
the heat and wave equati ons on an unbounded domai n 277
Intuitively, it seems apparent that w(x, t) = 0 for all real x and t > 0. To prove
this rigorously, we begin by multiplying both sides of w
t
= κw
xx
by w to get
ww
t
= κww
xx
.
Next, we make the sneaky observation that the lefthand side can be rewritten
as follows:
1
2
_
w
2
_
t
= κww
xx
.
Integrating both sides over the entire spatial domain yields
_
∞
−∞
1
2
_
w
2
_
t
dx = κ
_
∞
−∞
ww
xx
dx
and, rewriting the lefthand side,
1
2
d
dt
_
∞
−∞
w
2
dx = κ
_
∞
−∞
ww
xx
dx.
Integrating the righthand side by parts,
1
2
d
dt
_
∞
−∞
w
2
dx = κ
_
ww
x
¸
¸
¸
¸
∞
−∞
−
_
∞
−∞
w
2
x
dx
_
.
Because of our assumption that φ(x) → 0 as x → ±∞, we know that w → 0 as
x → ±∞. Therefore, the boundary term that appeared when integrating by parts
is actually zero:
1
2
d
dt
_
∞
−∞
w
2
dx = −κ
_
∞
−∞
w
2
x
dx.
The important thing to notice here is that the integrand on the righthand side is
nonnegative: w
2
x
≥ 0. Thus, we have an inequality
d
dt
_
∞
−∞
w
2
dx ≤ 0.
In words, this means that the area under the graph of w
2
(a nonnegative function)
cannot increase over time. Hence,
_
∞
−∞
w
2
(x, t) dx ≤
_
∞
−∞
w
2
(x, 0) dx
278 well posedness and the heat equati on
for all t > 0. However, we know (from the initial condition) that w
2
(x, 0) = 0,
which implies that
_
∞
−∞
w
2
(x, t) dx ≤ 0
for all t > 0. On the other hand, the integral could never be negative because
w
2
(x, t) ≥ 0 for all x and t. We have shown that
_
∞
−∞
w
2
(x, t) dx = 0,
which could only be true if w
2
= 0 throughout the entire domain −∞ < x < ∞.
Consequently, w = 0 so that u = v, and we see that the Cauchy problem has a
unique solution as claimed.
We can use almost identical computations to prove that the solution is not
greatly affected by small changes in the initial condition φ. To see this, let us
consider two Cauchy problems
u
t
−κu
xx
= 0 u(x, 0) = φ
1
(x)
and
v
t
−κv
xx
= 0 v(x, 0) = φ
2
(x).
Both u and v satisfy the heat equation but with different initial conditions. To
measure the impact of altering the initial conditions, we will compare the gap
between u(x, t) and v(x, t) with the gap between the initial conditions φ
1
and φ
2
.
Letting w = u −v as before, w satisﬁes the heat equation. However, this time the
initial condition on w is nonzero, namely
w(x, 0) = u(x, 0) −v(x, 0) = φ
1
(x) −φ
2
(x).
Retracing the same steps we performed above, we know that
_
∞
−∞
w
2
(x, t) dx ≤
_
∞
−∞
w
2
(x, 0) dx.
Equivalently,
_
∞
−∞
[u(x, t) −v(x, t)]
2
dx ≤
_
∞
−∞
[φ
1
(x) −φ
2
(x)]
2
dx
the heat and wave equati ons on an unbounded domai n 279
for all time t ≥ 0. The integral on the right hand side provides a measure of
the “gap” between φ
1
and φ
2
. Anytime the graphs φ
1
and φ
2
deviate from each
other, a positive contribution is made to the integral. Likewise, the integral on
the lefthand side of the inequality gives us a measure of the “gap” between
u(x, t) and v(x, t) at time t. The inequality assures us that the separation between
u and v can never increase over time. We conclude that small changes in the
initial conditions cannot cause major differences between how solutions behave,
and hence we have shown that the Cauchy problem for the heat equation is
wellposed. We remark that there are other (more natural) ways to measure the
separation between two functions, rather than integrating the square of their
difference. The above measure of “separation” between φ
1
(x) and φ
2
(x) is the
square of the L
2
distance between those functions, a concept that we will explore
in greater depth in subsequent chapters.
Dirichlet Problem. The energy method also works to establish wellposedness
of the Dirichlet problem for the heat equation:
u
t
= κu
xx
, (0 < x < L)
u(x, 0) = φ(x), (0 < x < L)
u(0, t) = g(t), (t > 0)
u(L, t) = h(t), (t > 0).
Existence of a solution will be established in a later chapter, once we develop the
separation of variables technique for pdes. Assuming that this Dirichlet problem
has a solution, we can show that it is unique and is not sensitive to small changes
in the initial condition or boundary conditions. Proceeding as above, suppose
that u and v are solutions of the Dirichlet problem. Deﬁning w = u −v, we see
that w satisﬁes the simpler problem
w
t
= κw
xx
, (0 < x < L)
w(x, 0) = 0, (0 < x < L)
w(0, t) = 0, (t > 0)
w(L, t) = 0, (t > 0).
280 well posedness and the heat equati on
The only difference in the procedure we used for the Cauchy problem is that we
integrate the equation
1
2
_
w
2
_
t
= κww
xx
over the ﬁnite spatial domain [0, L]:
_
L
0
1
2
_
w
2
_
t
dx = κ
_
L
0
ww
xx
dx.
Integrating the righthand side by parts,
1
2
d
dt
_
L
0
w
2
dx = κ
_
ww
x
¸
¸
¸
¸
L
0
−
_
L
0
w
2
x
dx
_
.
This time, the boundary term (ww
x
) vanishes because the boundary conditions
tell us that w = 0 at both x = 0 and x = L. The remainder of the calculation
is precisely the same as before, and we conclude that the Dirichlet problem for
the heat equation has a unique solution. Notice that unlike the proof of well
posedness for the Cauchy problem, we did not require any special assumptions
regarding the initial or boundary conditions. Proving that the solution of the
Dirichlet problem is not sensitive to small changes in the initial (or boundary)
conditions is essentially identical to the corresponding proof for the Cauchy
problem. Again, the only difference is that all integrals are evaluated over
the ﬁnite interval [0, L] as opposed to the inﬁnite domain (−∞, ∞). These
calculations, as well as the proof of wellposedness for the Neumann problem,
are left as exercises.
The Maximum Principle. We now use a special property of the heat equation
to provide a much quicker proof that the Dirichlet problem is wellposed. In the
process, we will gain even more intuition regarding how solutions of the heat
equation can behave.
Theorem 10.4.1 (Maximum Principle). Suppose that u(x, t) satisﬁes the heat
equation on the interval 0 ≤ x ≤ L and for t ≥ 0. Then the maximum value of
u(x, t) occurs either initially (i.e., when t = 0) or on the boundary (i.e., x = 0 or
x = L).
The intuitive explanation of the Maximum Principle as follows. If the maxi
mum temperature within a onedimensional wire occurred somewhere in the
the heat and wave equati ons on an unbounded domai n 281
middle of the wire, then it would have to cool off unless heat were supplied at
the boundary.
The idea behind the proof of the Maximum Principle involves some notions
from ﬁrstsemester calculus. Suppose indirectly that u has a maximum at
some point (x, t) in the interior of the domain. Then we would expect the ﬁrst
derivatives u
t
and u
x
to be zero at that point. Moreover, the second derivative
u
xx
should be at most zero. (After all, if u
xx
(x, t) > 0, then a plot of u versus
x would reveal a local minimum, not a maximum.) It cannot be the case that
u
xx
< 0, because this would imply that 0 = u
t
= κu
xx
< 0, a contradiction. We
have ruled out the possibility that u could have a maximum at a point (x, t) on
the interior of the domain if u
xx
< 0. The proof that no interior maximum could
occur with u
xx
= 0 requires more care, and the details appear in Strauss [10].
The same intuition that motivated the Maximum Principle gives rise to a
similar result:
Theorem 10.4.2 (Minimum Principle.). Suppose that u(x, t) satisﬁes the heat
equation on the interval 0 ≤ x ≤ L and for t ≥ 0. Then the minimum value of
u(x, t) occurs either initially (i.e., when t = 0) or on the boundary (i.e., x = 0 or
x = L).
Proof. If u(x, t) satisﬁes u
t
= κu
xx
then so does −u, since the heat equation is
linear and homogeneous. Applying the maximum principle to −u, we conclude
that the maximum value of −u must occur either initially (t = 0) or at one of the
boundaries (x = 0 or x = L). Certainly u attains its minimum value wherever
−u attains its maximum value.
The Maximum and Minimum Principles can be used to provide a quick proof
that the Dirichlet problem
u
t
= κu
xx
, (0 < x < L)
u(x, 0) = φ(x), (0 < x < L)
u(0, t) = g(t), (t > 0)
u(L, t) = h(t), (t > 0)
282 well posedness and the heat equati on
has a unique solution. As before, suppose that u and v are both solutions and
deﬁne w = u −v. Then w satisﬁes the simpler Dirichlet problem
w
t
= κw
xx
, (0 < x < L)
w(x, 0) = 0, (0 < x < L)
w(0, t) = 0, (t > 0)
w(L, t) = 0, (t > 0).
Notice that w is exactly 0 initially and at both boundaries. By the Maximum and
Minimum principles, the maximum and minimum values of w are both 0. This
means that w(x, t) = 0 for all (x, t), implying that u = v.
Testing for stability with respect to changes in the initial (or boundary) condi
tions is also straightforward using the Maximum/Minimum Principles. Suppose
that u and v represent solutions to the Dirichlet problems
u
t
= κu
xx
, (0 < x < L)
u(x, 0) = φ
1
(x), (0 < x < L)
u(0, t) = g(t), (t > 0)
u(L, t) = h(t), (t > 0)
and
v
t
= κv
xx
, (0 < x < L)
v(x, 0) = φ
2
(x), (0 < x < L)
v(0, t) = g(t), (t > 0)
v(L, t) = h(t), (t > 0).
The only difference between the two Dirichlet problems is in the initial condition.
Deﬁning w = u −v as usual, we see that w satisﬁes a new Dirichlet problem:
w
t
= κw
xx
, (0 < x < L)
w(x, 0) = φ
1
(x) −φ
2
(x), (0 < x < L)
w(0, t) = 0, (t > 0)
w(L, t) = 0, (t > 0).
the heat and wave equati ons on an unbounded domai n 283
We now apply the Maximum/Minimum Principles to w. Since w = 0 on the
boundaries, we have
max
0≤x≤L, 0≤t
w(x, t) ≤ max
0≤x≤L
φ
1
(x) −φ
2
(x).
Equivalently,
max
0≤x≤L, 0≤t
u(x, t) −v(x, t) ≤ max
0≤x≤L
φ
1
(x) −φ
2
(x).
This estimate shows that the Dirichlet problem is stable to small changes in
the initial condition. Speciﬁcally, it tells us that the maximum gap between
the solutions u and v of the two problems is at most as large as the maximum
separation between the two different initial conditions. Hence, if we start from
two different initial conditions φ
1
(x) and φ
2
(x) that are “close”, then the solutions
u(x, t) and v(x, t) will remain “close” for all time t. The same argument reveals
that the Dirichlet problem is not sensitive to small changes in the boundary
conditions at x = 0 and x = L.
We have now provided two different proofs that the Dirichlet problem for
the heat equation is wellposed: one using the energy method, and one using
the maximum principle. The proofs introduced two very different notions of
“separation”, or distance, between two functions φ
1
(x) and φ
2
(x). When we
proved stability via the energy method, we used the integral
_
L
0
[φ
1
(x) −φ
2
(x)]
2
dx
to measure the distance between the two functions. By contrast, the Maximum
Principle used a different measure of distance, namely
max
0≤x≤L
φ
1
(x) −φ
2
(x).
Both of these are perfectly reasonable ways of measuring the “distance” between
two functions, and our choice of which metric to use often depends upon the
context. Later, during our study of Fourier series, we will explore these different
notions of distance in greater detail.
284 i nhomogeneous equati ons and duhamel’ s pri nci ple
10.5. Inhomogeneous Equations and Duhamel’s Principle
Suppose that L is a linear differential operator, and consider the linear, inho
mogeneous differential equation L(u) = f . In your introductory course on
differential equations, you learned that the general solution of L(u) = f has the
form u = u
h
+ u
p
, where u
h
is the general solution of the homogeneous problem
L(u) = 0 and u
p
is a particular solution of the inhomogeneous equation L(u) = f .
In this section, we will solve the Cauchy problems for the inhomogeneous heat,
wave, and transport equations by using Duhamel’s Principle, a generalization
of the variation of parameters technique from ordinary differential equations.
For this reason, we motivate Duhamel’s Principle by performing an alternate
derivation of the variation of parameters formula (2.18).
Consider the ordinary differential equation
du
dt
= Au + f (t)
u(0) = u
0
,
(10.25)
where A is a constant and f is continuously differentiable (i.e., f has a continuous
derivative). By linearity, we can split this initial value problem into two simpler
problems, writing u = v + w where
dv
dt
= Av
v(0) = u
0
,
(10.26)
and
dw
dt
= Aw + f (t)
w(0) = 0.
(10.27)
The solution of the homogeneous problem for v is v(t) = e
tA
u
0
. When we solve
the inhomogeneous problem (10.27), we will invoke a technical Lemma:
Lemma 10.5.1. Suppose that f (x, t) and ∂ f /∂t are continuous and that β(t) is
differentiable. Then
d
dt
_
β(t)
0
f (x, t) dx = β
/
(t) f (β(t), t) +
_
β(t)
0
∂ f
∂t
(x, t) dx.
the heat and wave equati ons on an unbounded domai n 285
Proof. Use the chain rule and the fundamental theorem of calculus.
Notice that Lemma 10.5.1 is not merely a restatement of the fundamental
theorem of calculus, because the variable t appears both in the integrand and in
the upper limit of integration.
For our purposes, a special case of Lemma 10.5.1 in which β(t) = t will be
especially useful:
Corollary 10.5.2. Suppose that f (x, t; s) and ∂ f /∂t are continuous functions of
the variables x and t as well as the parameter s. Then
d
dt
_
t
0
f (x, t −s; s) ds = f (x, 0; s) +
_
t
0
∂ f
∂t
(x, t −s; s) ds.
We will use Corollary 10.5.2 to show that the solution of (10.27) can be con
structed by integrating the solution of a homogeneous initial value problem,
namely
dW
dt
= AW
W(0; s) = f (s),
(10.28)
where s is a parameter. Notice that the inhomogeneous term in (10.27) has been
transferred to the initial condition in (10.28).
Duhamel’s Principle: We claim that if W(t; s) is a solution of (10.28), then
w(t) =
_
t
0
W(t −s; s) ds (10.29)
is a solution of (10.27). To see why, suppose that w(t) is deﬁned as in (10.29).
Then by Corollary (10.5.2),
w
/
(t) =
d
dt
_
t
0
W(t −s; s) ds = W(0; t) +
_
t
0
W
t
(t −s; s) ds
= f (t) + A
_
t
0
W(t −s; s) ds = f (t) + Aw(t),
as claimed.
Since the solution of (10.28) is given by W(t; s) = e
tA
f (s), the above claim tells
us that
w(t) =
_
t
0
W(t −s; s) ds =
_
t
0
e
(t−s)A
f (s) ds
286 i nhomogeneous equati ons and duhamel’ s pri nci ple
is a solution of (10.27). Finally, the general solution of our original problem (10.25)
is given by
u(t) = v(t) + w(t) = e
tA
u
0
+
_
t
0
e
(t−s)A
f (s) ds, (10.30)
which we recognize as the familiar variation of parameters formula.
The same idea can be used to solve inhomogeneous pdes, which we now
demonstrate via three examples.
Inhomogeneous heat equation. Recall that the Cauchy problem for the [homo
geneous] heat equation models heat transfer within a long, thin wire. The wire
is insulated to prevent heat from radiating outward, and the diffusion of heat
is not inﬂuenced by any sources/sinks of heat. Now consider the more general
diffusion problem in which we are allowed to supply/remove heat from the wire,
and let f (x, t) describe the heat injected at position x and time t ≥ 0. (We will
assume that f is a continuously differentiable function.) Mathematically, we can
model this scenario with the Cauchy problem
u
t
= κu
xx
+ f (x, t)
u(x, 0) = φ(x)
(10.31)
where, as usual, φ(x) denotes the initial temperature distribution within the
wire.
We will solve the inhomogeneous Cauchy problem (10.31) via the same pro
cedure used to solve the ode (10.25). First, write u = v + w where v and w are
solutions of the simpler problems
v
t
= κv
xx
v(x, 0) = φ(x)
(10.32)
and
w
t
= κw
xx
+ f (x, t)
w(x, 0) = 0.
(10.33)
Because v satisﬁes the homogeneous Cauchy problem for the heat equation, we
immediately conclude that
the heat and wave equati ons on an unbounded domai n 287
v(x, t) = (S φ)(x, t) =
_
∞
−∞
S(x −y, t) φ(y) dy
=
1
√
4πκt
_
∞
−∞
e
−(x−y)
2
/4κt
φ(y) dy,
where S denotes the heat kernel (10.24). To solve (10.33), we mimic the preceding
example.
Duhamel’s Principle: We claim that if W(x, t; s) is a solution of the homogeneous
problem
W
t
= κW
xx
W(x, 0; s) = f (x, s),
(10.34)
where s is treated as a parameter, then
w(x, t) =
_
t
0
W(x, t −s; s) ds (10.35)
is a solution of (10.33). To prove the claim, suppose that w(x, t) is given by (10.35).
By computing the relevant partial derivatives of w(x, t), we must show that w
satisﬁes the inhomogeneous heat equation. According to Corollary 10.5.2,
w
t
(x, t) = W(x, 0; t) +
_
t
0
W
t
(x, t −s; s) ds = f (x, t) +
_
t
0
W
t
(x, t −s; s) ds,
and by direct differentiation,
w
xx
(x, t) =
_
t
0
W
xx
(x, t −s; s) ds.
Therefore,
w
t
−κw
xx
= f (x, t) +
_
t
0
W
t
(x, t −s; s) −κW
xx
(x, t −s; s) ds
= f (x, t) +
_
t
0
[W
t
−κW
xx
](x, t −s; s) ds = f (x, t),
as claimed.
Now since the solution of the homogeneous problem (10.34) is given by
W(x, t; s) = (S f )(x, t; s) =
_
∞
−∞
S(x −y, t) f (y, s) dy,
288 i nhomogeneous equati ons and duhamel’ s pri nci ple
it follows from Duhamel’s Principle that
w(x, t) =
_
t
0
W(x, t −s; s) ds =
_
t
0
_
∞
−∞
S(x −y, t −s) f (y, s) dy ds
is a solution of (10.33). Finally, recalling that u = v + w, we conclude that the
solution of the Cauchy problem (10.31) for the inhomogeneous heat equation is
given by
u(x, t) =
_
∞
−∞
S(x −y, t)φ(y) dy +
_
t
0
_
∞
−∞
S(x −y, t −s) f (y, s) dy ds (10.36)
for t > 0. From the deﬁnition of the heat kernel S(x, t), formula (10.36) is
equivalent to
u(x, t) =
1
√
4πκt
_
∞
−∞
e
−(x−y)
2
/4κt
φ(y) dy
+
_
t
0
_
∞
−∞
1
_
4πκ(t −s)
e
−(x−y)
2
/4κ(t−s)
f (y, s) dy ds.
The solution (10.36) is rather unwieldy, as there are very few choices of initial
heat proﬁles φ(x) and heat source terms f (x, t) for which these integrals can be
evaluated explicitly. Hence, one typically uses computer software packages to
numerically approximate solutions of inhomogeneous pdes such as this one.
Inhomogeneous wave equation. In the preceding examples, we used Duhamel’s
Principle to solve inhomogeneous equations whose homogeneous counterparts
had already been solved. In effect, Duhamel’s Principle tells us that the cumula
tive effect of a source term f (t) or f (x, t) is measured by taking an appropriate
integral involving (i) a solution of the underlying homogeneous problem and (ii)
the function f itself. We apply the same concept to solve the Cauchy problem
for the inhomogeneous wave equation:
u
tt
= c
2
u
xx
+ f (x, t)
u(x, 0) = φ(x)
u
t
(x, 0) = ψ(x).
(10.37)
Here, φ(x) and ψ(x) represent the initial displacement and velocity (respectively)
of a vibrating string, and f (x, t) describes an external force being applied to
the heat and wave equati ons on an unbounded domai n 289
the string. We will assume that f (x, t) is a continuously differentiable function.
Splitting (10.37) into two simpler problems, we write u = v + w where
v
tt
= c
2
v
xx
v(x, 0) = φ(x)
v
t
(x, 0) = ψ(x).
(10.38)
and
w
tt
= c
2
w
xx
+ f (x, t)
w(x, 0) = 0
w
t
(x, 0) = 0.
(10.39)
Because v satisﬁes the homogeneous Cauchy problem for the wave equation, we
quote D’Alembert’s formula to obtain
v(x, t) =
1
2
[φ(x + ct) + φ(x −ct)] +
1
2c
_
x+ct
x−ct
ψ(s) ds.
To solve (10.39), we proceed as before.
Duhamel’s Principle: We claim that if W(x, t; s) is a solution of the homogeneous
problem
W
tt
= c
2
W
xx
W(x, 0; s) = 0
W
t
(x, 0; s) = f (x, s),
(10.40)
where s is treated as a parameter, then
w(x, t) =
_
t
0
W(x, t −s; s) ds (10.41)
is a solution of (10.39). To prove the claim, suppose that w(x, t) is given by (10.41).
By computing the relevant partial derivatives of w(x, t), we must show that w
satisﬁes the inhomogeneous wave equation. According to Corollary 10.5.2,
w
t
(x, t) = W(x, 0; t) +
_
t
0
W
t
(x, t −s; s) ds =
_
t
0
W
t
(x, t −s; s) ds
290 i nhomogeneous equati ons and duhamel’ s pri nci ple
and, differentiating with respect to t a second time,
w
tt
(x, t) = W
t
(x, 0; t) +
_
t
0
W
tt
(x, t −s; s) ds = f (x, t) +
_
t
0
W
tt
(x, t −s; s) ds.
Derivatives with respect to x can be computed directly:
w
xx
(x, t) =
_
t
0
W
xx
(x, t −s; s) ds.
Therefore,
w
tt
−c
2
w
xx
= f (x, t) +
_
t
0
W
tt
(x, t −s; s) −c
2
W
xx
(x, t −s; s) ds
= f (x, t) +
_
t
0
[W
tt
−c
2
W
xx
](x, t −s; s) ds = f (x, t),
as claimed.
From D’Alembert’s formula, the solution of the homogeneous problem (10.40)
is given by
W(x, t; s) =
1
2c
_
x+ct
x−ct
f (η, s) dη,
from which it follows that
w(x, t) =
_
t
0
W(x, t −s; s) ds =
1
2c
_
t
0
_
x+c(t−s)
x−c(t−s)
f (η, s) dη ds
is a solution of (10.39). Finally, recalling that u = v + w, we conclude that the
solution of the Cauchy problem (10.37) for the inhomogeneous wave equation is
given by
u(x, t) =
1
2
[φ(x + ct) + φ(x −ct)] +
1
2c
_
x+ct
x−ct
ψ(s) ds
+
1
2c
_
t
0
_
x+c(t−s)
x−c(t−s)
f (η, s) dη ds.
(10.42)
The iterated integral appearing in (10.42) is not nearly as problematic as the one
that appeared in the solution of the inhomogeneous heat equation (see exercises).
the heat and wave equati ons on an unbounded domai n 291
Inhomogeneous transport equation. In a previous chapter, we showed that the
solution of the Cauchy problem
u
t
+ cu
x
= 0
u(x, 0) = f (x)
(10.43)
for the simple transport equation is given by u(x, t) = f (x − ct). Recall that
u(x, t) can be interpreted as the concentration of a pollutant being carried in a
river whose current has constant velocity c. The pollutant does not diffuse (it
is merely suspended in the water) and the total mass of pollutant is conserved.
The inhomogeneous transport equation
u
t
+ cu
x
= g(x, t)
u(x, 0) = f (x),
(10.44)
allows for the possibility that we may add/remove pollutant from the stream, as
modeled by the source term g(x, t). Solving (10.44) follows the same procedure
that we used to solve the inhomogeneous heat and wave equations. As an
exercise, use Duhamel’s Principle to show that the solution of (10.44) is given by
u(x, t) = f (x −ct) +
_
t
0
g(x −c(t −s), s) ds. (10.45)
Exercises
1. Solve the pde u
tt
−5u
xt
−36u
xx
= 0. To do so, factor the operator on u as
_
∂
∂t
−9
∂
∂x
__
∂
∂t
+4
∂
∂x
_
u = 0,
and mimic what we did when solving the wave equation.
2. Solve the pde u
tt
− (α + β)u
xt
+ αβu
xx
= 0, where α and β are nonzero
constants. (Hint: See previous exercise.)
3. Let c ,= 0 be a constant. Solve the Cauchy problem
u
tt
−c
2
u
xx
= 0
u(x, 0) = e
−x
292 i nhomogeneous equati ons and duhamel’ s pri nci ple
u
t
(x, 0) = cos x.
4. Solve the Cauchy problem
u
tt
−9u
xx
= 0
u(x, 0) = arctan(x)
u
t
(x, 0) = 2x ln(1 + x
2
).
5. A function f (x) is called even if it satisﬁes f (−x) = f (x) for all real x. For
example, cos(x) is even because cos(−x) = cos(x) for all x. Show that if the
initial conditions φ(x) and ψ(x) for the wave equation are both even, then
the solution u(x, t) is even (in the x variable) for all time t. That is, show that
u(−x, t) = u(x, t).
6. Suppose that u(x, t) satisﬁes the wave equation u
tt
= c
2
u
xx
. Let α be any
nonzero constant and deﬁne the dilated function v(x, t) = u(αx, αt). Show
that v also satisﬁes the wave equation.
7. Solve the Cauchy problem
u
tt
−c
2
u
xx
= 0 (−∞ < x < ∞),
u(x, 0) = sin x,
u
t
(x, 0) = −sin x.
Express the solution as a product of two functions, one of which depends
on x and one of which depends on t. To do so, you may ﬁnd the following
trigonometric identities useful:
sin(α +β) = sin α cos β +cos α sin β, cos(α +β) = cos α cos β −sin α sin β.
8. Consider the Cauchy problem
u
t
= κu
xx
(−∞ < x < ∞),
u(x, 0) = φ(x),
where φ(x) is an even function. Show that the solution u will remain even in
x for all t > 0. In other words, show that u(−x, t) = u(x, t) for all t > 0.
the heat and wave equati ons on an unbounded domai n 293
9. The function u(x, t) = 10 − x + 4t + 2x
2
is a solution of the heat equation
u
t
= u
xx
. Find the locations of its maximum and minimum values on the
domain 0 ≤ x ≤ 2 and 0 ≤ t ≤ 8.
10. Show that u(x, t) = e
αx+βt
satisﬁes the heat equation u
t
= κu
xx
if and only if
β = κα
2
. In the special case α = −2 and κ = 1, note that e
−2x+4t
is a solution
of u
t
= u
xx
. Find the maximum and minimum values of e
−2x+4t
over the
domain 0 ≤ x ≤ 2 and 0 ≤ t ≤ 3. Do the Maximum/Minimum Principles
hold?
11. Unlike the heat equation, the wave equation u
tt
= u
xx
does not obey a
Maximum Principle. Consider, for example, the function u(x, t) = sin(x +
t) + cos(x −t) on the domain 0 ≤ x ≤
π
2
and 0 ≤ t ≤
π
4
. Show that u(x, t)
satisﬁes the wave equation and ﬁnd the maximum value of u(x, t) over the
given domain. Then, explain why you can conclude that the wave equation
does not obey the Maximum Principle.
12. Suppose that u(x, t) and v(x, t) are solutions of the heat equation on the
domain 0 ≤ x ≤ L and t ≥ 0. Suppose that u ≤ v both initially (t = 0) and at
both boundaries (x = 0 and x = L). Use the Maximum/Minimum Principle
to prove that u(x, t) ≤ v(x, t) for all (x, t). Hint: Let w = v −u.
13. According to Lemma 10.3.1, we know that
_
∞
−∞
e
−r
2
dr =
√
π.
By making the substitution r = x/
√
4κt, show that
_
∞
−∞
S(x, t) dx = 1 (t > 0),
where S(x, t) denotes the heat kernel.
14. Consider the Dirichlet problem
u
t
= κu
xx
(0 < x < L and t > 0)
u(x, 0) = φ(x) (0 < x < L)
u(0, t) = g(t)
u(L, t) = h(t).
294 i nhomogeneous equati ons and duhamel’ s pri nci ple
Show that the solution u(x, t) is not sensitive to small changes in the right
boundary condition h(t).
15. Let r > 0 be a constant. The reactiondiffusion equation
u
t
−κu
xx
+ ru = 0 (−∞ < x < ∞),
u(x, 0) = φ(x),
can be solved by making a substitution that converts it to the heat equation.
(a) Let v(x, t) = e
rt
u(x, t). Show that v satisﬁes the heat equation v
t
= κv
xx
,
and obeys the same initial condition as u. (b) After quoting (10.23) to write
down a formula for v(x, t), ﬁnd the solution u(x, t) of the original reaction
diffusion problem.
16. Let c > 0 be a constant. The advectiondiffusion equation
u
t
+ cu
x
= κu
xx
(−∞ < x < ∞),
u(x, 0) = φ(x)
models concentration of a pollutant that diffuses while being transported with
constant speed c. In the absence of diffusion (i.e., κ = 0), the pde reduces to
the simple transport equation which was solved in a previous chapter. From
our understanding of the transport equation, the reactiondiffusion equation
should simplify if we switch to a moving coordinate system. (a) Make the
substitution v(x, t) = u(x + ct, t) and show that v satisﬁes the heat equation.
(b) Set up and solve a Cauchy problem for v(x, t). (c) Use your formula from
Part (b) to write down the solution of the advectiondiffusion equation.
17. Consider the Neumann problem for the heat equation:
u
t
= κu
xx
(0 < x < L)
u(x, 0) = φ(x) (0 < x < L)
u
x
(0, t) = g(t)
u
x
(L, t) = h(t).
The purpose of this problem is to show that, assuming a solution exists, it
must be unique. (i) Explain why we cannot use the Maximum/Minimum
the heat and wave equati ons on an unbounded domai n 295
Principle to prove uniqueness as we did for the Dirichlet problem. (ii) Use the
energy method to prove that the solution really is unique.
18. The purpose of this exercise is to use the energy method to prove that if the
Dirichlet problem
u
tt
= c
2
u
xx
(0 < x < L)
u(x, 0) = φ(x) (0 < x < L)
u
t
(x, 0) = ψ(x) (0 < x < L)
u(0, t) = g(t)
u(L, t) = h(t)
for the wave equation has a solution, then the solution must be unique.
(a) Suppose that u and v are solutions of this Dirichlet problem. Show that
w = u −v satisﬁes the wave equation with w = w
t
= 0 initially and w = 0
on the boundaries.
(b) Deﬁne the energy function
E(t) =
1
2
_
L
0
w
2
t
+ c
2
w
2
x
dx.
Assuming that you are justiﬁed in differentiating under the integral sign,
show that
E
/
(t) =
_
L
0
w
t
w
tt
+ c
2
w
x
w
xt
dx =
_
L
0
w
t
w
tt
dx + c
2
_
L
0
w
x
w
xt
dx.
(c) Integrate w
x
w
xt
by parts and show that
E
/
(t) =
_
L
0
w
t
_
w
tt
−c
2
w
xx
_
dx = 0.
(d) Use the result from Part (c) to explain why E(t) = 0 for all t. Then,
explain why it follows that w(x, t) = 0 for 0 ≤ x ≤ L and all t. Finally,
conclude that u(x, t) = v(x, t), which means that the solution of the
Dirichlet problem is unique.
19. Use Duhamel’s Principle to derive formula (10.45), the solution of the inho
mogeneous transport equation (10.44).
296 i nhomogeneous equati ons and duhamel’ s pri nci ple
20. Solve the Cauchy problem
u
t
+8u
x
= x sin t
u(x, 0) =
1
1 + x
2
.
21. Solve the Cauchy problem
u
tt
= c
2
u
xx
+ xt
2
u(x, 0) = e
−x
2
u
t
(x, 0) = 0.
How does the source term xt
2
affect the behavior of the solution? That is,
how does your solution compare with that of the associated homogeneous
problem?
22. Solve the Cauchy problem
u
tt
= 4u
xx
+ e
x
u(x, 0) = 0
u
t
(x, 0) = 0
and sketch the solution u(x, 1) versus x. Note: If the source term e
x
were
absent, the solution would be u(x, t) = 0. Physically, this would correspond
to a stationary string in its equilibrium position. What effect does the source
term have on the displacement of the string?
CHAPTER 11
InitialBoundary Value Problems
T
he inﬁnite spatial domains considered in the previous chapter give insight
regarding the behavior of waves and diffusions. However, since such
domains are not physically realistic, we need to develop new techniques for
solving pdes on bounded domains. As a ﬁrst step towards solving the heat and
wave equations over ﬁnite spatial domains (such as the interval 0 ≤ x ≤ L in
one space dimension), we will solve these equations on “semiinﬁnite” domains
whose boundaries consist of one point.
11.1. Heat and Wave Equations on a HalfLine
We begin by solving the homogeneous Dirichlet problem for the heat equation
on the interval 0 ≤ x < ∞; that is,
u
t
= κu
xx
(0 < x < ∞) (11.1)
u(x, 0) = φ(x) (0 < x < ∞) (11.2)
u(0, t) = 0 (t ≥ 0). (11.3)
The homogeneous boundary condition is quite important for the solution tech
nique that follows. In the context of heat transfer within a “onedimensional”
wire, this Dirichlet boundary condition is analogous to immersing the x = 0 end
of the wire in a bath of ice water with temperature zero degrees Celsius.
We will solve the homogeneous Dirichlet problem (11.1)–(11.3) using a reﬂection
method, temporarily extending our spatial domain to the entire real line and
solving a Cauchy problem instead. By quoting the formula for the solution
297
298 heat and wave equati ons on a half li ne
φ
odd
(x) φ(x)
x x
Figure 11.1. Illustration of the odd extension of a function φ(x).
of the Cauchy problem in the preceding chapter, we will obtain the solution
of (11.1)–(11.3) by restricting ourselves to the original spatial domain.
First, recall that a function f (x) of a single variable is called odd if it has the
property that f (−x) = −f (x) for all real x. Examples of odd functions include
sin(x) and x
3
. If f is an odd function, notice that f (0) = f (−0) = −f (0), which
implies that f (0) = 0. Now, referring to the initial condition (11.2) above, we
deﬁne the odd extension of φ(x) as
φ
odd
(x) =
_
¸
¸
¸
_
¸
¸
¸
_
φ(x) if x > 0
−φ(−x) if x < 0
0 if x = 0.
By construction, φ
odd
is an odd function and is deﬁned for all real x (see Fig
ure 11.1). Now consider the Cauchy problem
v
t
= κv
xx
(−∞ < x < ∞)
v(x, 0) = φ
odd
(x) (−∞ < x < ∞).
From the previous chapter, we know that the solution is given by the convolution
of the heat kernel S(x, t) with the initial condition:
v(x, t) = (S φ
odd
)(x, t) =
_
∞
−∞
S(x −y, t)φ
odd
(y) dy.
We claim that the restriction of v(x, t) to the domain x ≥ 0 is the solution of
the Dirichlet problem (11.1)–(11.3). To see why, we need to verify that all three
conditions of our Dirichlet problem are satisﬁed. Certainly v(x, t) satisﬁes the
i ni ti al boundary value problems 299
same pde as u(x, t) on the domain x > 0. The initial conditions also match on
that domain, because v(x, 0) = φ(x) = u(x, 0) whenever x > 0. Checking the
boundary condition requires a bit more care. As an exercise, you should verify
that since the initial condition for v(x, t) is odd, then the solution v(x, t) will
remain odd for all t > 0. That is, v(−x, t) = −v(x, t) for all t ≥ 0. By our earlier
remarks on odd functions, this implies that v(0, t) = 0 for all t ≥ 0. It follows
that v automatically obeys the homogeneous Dirichlet boundary condition that
we imposed on u. Since u(x, t) and v(x, t) satisfy the heat equation with the same
initial and boundary conditions on the domain x ≥ 0 and t ≥ 0, we conclude (by
uniqueness) that u(x, t) = v(x, t) on that domain.
Now that we have proved that the restriction of v(x, t) to the domain x ≥ 0
is the solution of the Dirichlet problem (11.1)–(11.3), we can give an explicit
formula. The piecewise deﬁnition of φ
odd
suggests that we split the region of
integration as
v(x, t) =
_
∞
0
S(x −y, t)φ
odd
(y) dy +
_
0
−∞
S(x −y, t)φ
odd
(y) dy.
Using the deﬁnition of φ
odd
,
v(x, t) =
_
∞
0
S(x −y, t)φ(y) dy −
_
0
−∞
S(x −y, t)φ(−y) dy,
and substituting w = −y in the second integral yields
v(x, t) =
_
∞
0
S(x −y, t)φ(y) dy +
_
0
∞
S(x + w, t)φ(w) dw.
Reversing the limits of integration in the second integral, we have
v(x, t) =
_
∞
0
S(x −y, t)φ(y) dy −
_
∞
0
S(x + w, t)φ(w) dw.
Since w is simply a dummy variable of integration, we may revert to using y
instead. Combining the two integrals,
v(x, t) =
_
∞
0
[S(x −y, t) −S(x + y, t)]φ(y) dy.
300 heat and wave equati ons on a half li ne
Finally, writing out the heat kernel S explicitly, we have shown that the solution
of the Dirichlet problem (11.1)–(11.3) is given by
u(x, t) =
1
√
4πκt
_
∞
0
_
e
−(x−y)
2
/4κt
− e
−(x+y)
2
/4κt
_
φ(y) dy (11.4)
for t > 0. Note that the integral is taken over the entire spatial domain.
In deriving (11.4), it was very important that the boundary condition at
x = 0 was homogeneous. Otherwise, the solution v(x, t) of the auxiliary Cauchy
problem would not have automatically satisﬁed the boundary condition (11.3)
for u(x, t). Fortunately, given a more general Dirichlet condition u(0, t) = g(t),
it is still possible to use the odd reﬂection technique. The idea is to make a
special substitution w(x, t) = u(x, t) − g(t) that converts the inhomogeneous
boundary condition into a homogeneous one. It turns out that w then satisﬁes an
inhomogeneous heat equation which can be solved via yet another substitution
(see exercises).
Example 11.1.1. To understand the effect of the homogeneous Dirichlet boundary
condition, let us compare the solution (10.23) of the Cauchy problem for the heat
equation with formula (11.4) for a special choice of initial condition. Namely, if
we use the constant heat distribution φ(x) = 1, then the solution of the Cauchy
problem would be
u(x, t) = (S φ)(x, t) =
_
∞
−∞
S(x −y, t) dy,
where S denotes the heat kernel. From a past exercise, we know that the integral
of the heat kernel over the entire real line is precisely 1 for all t > 0. Therefore,
the solution of the Cauchy problem with this initial condition is u(x, t) = 1
for all real x and t ≥ 0. This makes perfect sense, because we would expect a
perfectly uniform initial heat distribution to remain uniform for all time t.
Now suppose we restrict ourselves to a halfline and impose a homogeneous
Dirichlet condition at x = 0:
u
t
= κu
xx
(0 < x < ∞)
u(x, 0) = 1 (0 < x < ∞)
u(0, t) = 0 (t ≥ 0).
i ni ti al boundary value problems 301
1
0
0 5
x
u
(
x
,
t
)
t = 0.0
t = 0.5
t = 2.0
Figure 11.2. Solution of (11.1)–(11.3) with constant initial heat distribu
tion φ(x) = 1.
By formula (11.4), the solution is
u(x, t) =
_
∞
0
S(x −y, t) − S(x + y, t) dy. (11.5)
Unlike the solution of the Cauchy problem, this integral is not identically equal
to 1 for all x ≥ 0 and t > 0. In fact, the integral (11.5) cannot be evaluated
explicitly (although it can be written in terms of the standard error function). A
graph of the function u(x, t) from (11.5) is shown in Figure 11.2. The “cooling”
effect of the boundary condition is felt over a progressively wider region near
the boundary.
Heat Equation, Neumann Problem The homogeneous Neumann problem for
the heat equation on a halfline is given by
u
t
= κu
xx
(0 < x < ∞) (11.6)
u(x, 0) = φ(x) (0 < x < ∞) (11.7)
u
x
(0, t) = 0 (t ≥ 0). (11.8)
The boundary condition (11.8) states that the spatial gradient of u is 0 at x = 0.
Hence, there is no heat ﬂux across the boundary, which is analogous to insulating
the x = 0 end of the wire to prevent “leakage” of heat. Physically, this is very
different from the homogeneous Dirichlet condition that we considered earlier.
302 heat and wave equati ons on a half li ne
φ(x) φ
even
(x)
x x
Figure 11.3. Illustration of the even extension of a function φ(x).
Solving (11.6)–(11.8) is accomplished by a similar reﬂection method as the one
we developed for the Dirichlet problem. However, using an odd extension of
φ(x) will not necessarily preserve the boundary condition at x = 0. Instead, we
will introduce the even extension of φ(x), which is deﬁned as
φ
even
(x) =
_
_
_
φ(x) if x ≥ 0
φ(−x) if x < 0.
This concept is illustrated in Figure 11.3. Notice that if φ(x) has a righthand
derivative of 0 at x = 0, then the even extension is differentiable at x = 0
and satisﬁes φ
/
even
(0) = 0. To solve our Neumann problem, follow a similar
procedure as before: (1) Solve the Cauchy problem using φ
even
as the initial
condition. (2) Argue that the solution of the Cauchy problem must remain even
for all t > 0 since the initial condition was even. (3) Show that restricting the
solution of the Cauchy problem to the domain x ≥ 0 yields the solution of the
Neumann problem. By following this procedure, you will ﬁnd that the solution
of (11.6)–(11.8) is given by
u(x, t) =
_
∞
0
[S(x −y, t) + S(x + y, t)] φ(y) dy
=
1
√
4πκt
_
∞
0
_
e
−(x−y)
2
/4κt
+ e
−(x+y)
2
/4κt
_
φ(y) dy
(11.9)
Notice that the only difference between the solutions (11.4) (Dirichlet problem)
and (11.9) (Neumann problem) is the sign that appears in the integrand.
Again, we remark that the method of even reﬂection relied upon the fact
that the Neumann boundary condition (11.8) was homogeneous. Given the
i ni ti al boundary value problems 303
more general Neumann condition u
x
(0, t) = g(t), we can make a substitution
w(x, t) = u(x, t) −xg(t). Then w(x, t) will satisfy the homogeneous Neumann
condition w
x
(0, t) = 0. Although this substitution improved the boundary
condition, it has the unfortunate sideeffect that w satisﬁes an inhomogeneous heat
equation. You will learn how to solve the inhomogeneous Neumann problem as
an exercise.
Wave Equation, Dirichlet Problem. The homogeneous Dirichlet problem for
the wave equation on a halfline is given by
u
tt
= c
2
u
xx
(0 < x < ∞) (11.10)
u(x, 0) = φ(x) (0 < x < ∞) (11.11)
u
t
(x, 0) = ψ(x) (0 < x < ∞) (11.12)
u(0, t) = 0 (t ≥ 0). (11.13)
Recall that the two initial conditions correspond to the initial displacement and
velocity of a vibrating string. The Dirichlet condition requires that the boundary
of the string remain stationary in the equilibrium position.
As with the homogeneous Dirichlet problem for the heat equation, we will use
the method of odd extension. Deﬁne
φ
odd
(x) =
_
¸
¸
¸
_
¸
¸
¸
_
φ(x) if x > 0
−φ(−x) if x < 0
0 if x = 0
and ψ
odd
(x) =
_
¸
¸
¸
_
¸
¸
¸
_
ψ(x) if x > 0
−ψ(−x) if x < 0
0 if x = 0
and consider the Cauchy problem
v
tt
= c
2
v
xx
(−∞ < x < ∞)
v(x, 0) = φ
odd
(x) (−∞ < x < ∞)
v
t
(x, 0) = ψ
odd
(x) (−∞ < x < ∞).
Once again, we claim that the restriction of v(x, t) to the domain x ≥ 0 is the
solution of the Dirichlet problem (11.10)–(11.13). Certainly u and v satisfy the
same pde and initial conditions on the domain x > 0, so it remains to check that
v automatically satisﬁes the boundary condition (11.13) on u. However, we know
that v will remain odd for all t because the initial conditions φ
odd
and ψ
odd
are
304 heat and wave equati ons on a half li ne
odd functions (exercise). This implies that v(0, t) = 0 for all t, which is precisely
what we need at the boundary.
The solution of the Cauchy problem is provided by D’Alembert’s formula:
v(x, t) =
1
2
[φ
odd
(x + ct) + φ
odd
(x −ct)] +
1
2c
_
x+ct
x−ct
ψ
odd
(s) ds. (11.14)
In order to recover the solution of (11.10)–(11.13), we must restrict this solution
to x ≥ 0 and express the solution in terms of φ and ψ (rather than φ
odd
and ψ
odd
).
We must exercise caution—even if x > 0 and t > 0, it is possible that x −ct < 0.
This would mean that the integral term in (11.14) is referencing points that lie
outside the spatial domain, a situation that we need to avoid. We will proceed
by considering several cases.
Case 1. Suppose that x −ct > 0 and x + ct > 0. Then the domain of integration
in formula (11.14) lies entirely within the spatial domain of our Dirichlet problem.
Moreover, we know that φ
odd
(x) = φ(x) and ψ
odd
(x) = ψ(x) for x > 0. Hence,
the solution of the Dirichlet problem is given by
u(x, t) =
1
2
[φ(x + ct) + φ(x −ct)] +
1
2c
_
x+ct
x−ct
ψ(s) ds. (11.15)
Notice that, in this case, formula (11.15) is identical to D’Alembert’s formula
for the solution of the Cauchy problem. This is explained easily if we recall the
notion of the domain of dependence for u(x, t). Since we have assumed that both
x −ct and x +ct are positive, the boundary x = 0 has no inﬂuence on u(x, t). This
is a consequence of the ﬁnite propagation speed c, as illustrated in the left panel
of Figure 11.4. Assuming t > 0 as illustrated in the Figure, the value of u(x, t) is
only impacted by the initial conditions within the interval [x −ct, x + ct]. Since
the interval avoids the boundary in this case, the boundary condition cannot
possibly inﬂuence the behavior of u at the point (x, t).
Case 2. Suppose that x − ct < 0 and x + ct > 0. This time, formula (11.14)
references points that lie outside the spatial domain (x ≥ 0) of our Dirichlet
problem. Writing the solution in terms of φ and ψ requires more care in this case.
Since x − ct < 0, we can write φ
odd
(x − ct) = φ
odd
(−(ct − x)) = −φ(ct − x).
Next, we must write the integral term in (11.14) in such a way that the interval
of integration does not include negative values. The idea is to split the integral
i ni ti al boundary value problems 305
x
t
x
t
(ct−x,0)
(x,t)
(x−ct,0) (x+ct,0) (x−ct,0) (x+ct,0)
(x,t)
Figure 11.4. Examples of domains of dependence for u(x, t), the solu
tion of the homogeneous Dirichlet problem (11.10)–(11.13). Left panel:
Domain of dependence of (x, t) assuming that t > 0 and both x ±ct > 0.
Right panel: Domain of dependence of (x, t) assuming that t > 0 and
x −ct < 0 < x + ct.
as follows:
1
2c
_
x+ct
x−ct
ψ
odd
(s) ds =
1
2c
_
0
x−ct
ψ
odd
(s) ds +
1
2c
_
x+ct
0
ψ
odd
(s) ds.
Referring to the deﬁnition of ψ
odd
, we can rewrite this equation as
1
2c
_
x+ct
x−ct
ψ
odd
(s) ds =
1
2c
_
0
x−ct
−ψ(−s) ds +
1
2c
_
x+ct
0
ψ(s) ds.
Substituting w = −s gives
1
2c
_
x+ct
x−ct
ψ
odd
(s) ds =
1
2c
_
0
ct−x
ψ(w) dw +
1
2c
_
x+ct
0
ψ(s) ds,
and since s and w are merely dummy variables of integration, we can now
combine the integrals as
1
2c
_
x+ct
x−ct
ψ
odd
(s) ds =
1
2c
_
x+ct
ct−x
ψ(s) ds.
In summary, we have now shown that if x −ct < 0 < x + ct, then the solution of
the Dirichlet problem (11.10)–(11.13) is
u(x, t) =
1
2
[φ(x + ct) −φ(ct −x)] +
1
2c
_
x+ct
ct−x
ψ(s) ds. (11.16)
306 separati on of vari ables
Formula (11.16) differs from D’Alembert’s formula. Interaction with the bound
ary forced us to “reﬂect” the negative quantity x −ct, expressing the solution
in terms of the positive quantity ct − x. This reﬂection off the boundary is
illustrated in the right panel of Figure 11.4. The shaded region shows the domain
of dependence for a point (x, t) with t > 0.
It is certainly possible to consider other possibilities, such as x ± ct < 0.
However, given that our spatial domain is x ≥ 0 and c > 0, such situations could
only occur for negative time t. Although it is perfectly reasonable to solve the
wave equation for negative t (contrast this with the heat equation), we will leave
the details of the x ± ct < 0 case to the reader.
Wave Equation, Neumann Problem. The homogeneous Neumann problem for
the wave equation on a halfline is given by
u
tt
= c
2
u
xx
(0 < x < ∞)
u(x, 0) = φ(x) (0 < x < ∞)
u
t
(x, 0) = ψ(x) (0 < x < ∞)
u
x
(0, t) = 0 (t ≥ 0).
Physically, the boundary condition is a bit more difﬁcult to interpret than a
homogeneous Dirichlet condition. Instead of holding the x = 0 end of our
vibrating string completely stationary, it is free to move transversely. Moreover,
there is no tension at that end of the string. To solve this Neumann problem,
deﬁne the even extensions of φ and ψ and convert it to a Cauchy problem as
before. Writing the solution involves several cases according to the signs of x +ct
and x −ct.
11.2. Separation of Variables
Henceforth, we will solve pdes on ﬁnite spatial domains. Since we generally work
in one spatial dimension, we have in mind a ﬁnite interval 0 ≤ x ≤ L, where L is
a positive constant. The boundary of the domain consists of two points (x = 0
and x = L), and we shall impose boundary conditions at both of these points.
In the context of the wave equation, L is the length of our vibrating string. For
i ni ti al boundary value problems 307
the heat equation, L would represent the length of the wire within which heat
diffuses longitudinally.
The inclusion of a second boundary point prevents us from recycling the
reﬂection methods introduced in the previous section—those methods only
apply to semiinﬁnite domains with a single boundary point. Instead, we will
develop a completely different approach known as separation of variables. To
motivate this technique, let us recall a speciﬁc example discussed in a previous
chapter. After deriving D’Alembert’s formula, we solved the Cauchy problem
u
tt
− c
2
u
xx
= 0 with initial displacement φ(x) = cos(x) and initial velocity
ψ(x) = −cos(x). By creative use of trigonometric identities, we were able to
express the solution in the form
u(x, t) = cos(x)
_
cos(ct) −
1
c
sin(ct)
_
.
The solution is written as a product of two functions: a function of x only
and a function of t only. The fact that we were able to “separate” the spatial
and temporal parts of the solution made it considerably easier to visualize the
dynamics.
The idea of the separation of variables technique is to seek special separated
solutions of the form u(x, t) = X(x)T(t), where X and T are functions only of
the spatial and temporal variables, respectively. By ﬁnding enough separated
solutions of a linear homogeneous initialboundary value problem, we will be
able to construct the general solution via superposition.
11.2.1 Wave Equation, Dirichlet Problem. We illustrate the method of separa
tion of variables by solving the homogeneous Dirichlet problem
u
tt
= c
2
u
xx
(0 < x < L) (11.17)
u(x, 0) = φ(x) (0 < x < L) (11.18)
u
t
(x, 0) = ψ(x) (0 < x < L) (11.19)
u(0, t) = 0 (t ≥ 0) (11.20)
u(L, t) = 0 (t ≥ 0). (11.21)
308 separati on of vari ables
Physically, the two Dirichlet boundary conditions indicate that the ends of our
vibrating string must remain stationary for all t, which is precisely what we
would envision for a vibrating guitar string.
Step 1: Separate the variables. The ﬁrst step in attacking this problem is to seek
nontrivial
1
separated solutions u(x, t) = X(x)T(t) which satisfy the pde (11.17)
and both boundary conditions (11.20)–(11.21). Substituting u(x, t) = X(x)T(t)
into the wave equation (11.17), we obtain
X(x)T
//
(t) = c
2
X
//
(x)T(t).
We are now able to use primes to denote differentiation, since X and T are
each functions of one variable. By algebra, we may separate the time and
spacedependent parts of this equation:
X
//
X
=
T
//
c
2
T
.
Notice that the lefthand side is a function of x only, whereas the righthand side
is a function of t only. The only way for such functions to be identically equal is
if both functions are equal to a common constant. Denoting this constant by −λ,
where the negative sign is included for later convenience, we have
X
//
X
=
T
//
c
2
T
= −λ.
Equivalently, we have obtained a pair of two secondorder, linear constant
coefﬁcient odes:
X
//
+ λX = 0 and T
//
+ λc
2
T = 0. (11.22)
The fact that we are now dealing with odes echoes a recurring theme: seeking
special solutions of pdes often reduces the pde to an ode.
Step 2: Boundary conditions. The boundary conditions (11.20)–(11.21) will
impose special requirements on the X equation in (11.22). Combining u(x, t) =
X(x)T(t) with the fact that u(0, t) = 0 for all time t, it follows that X(0)T(t) = 0
for all time t. There are two ways this could happen: either X(0) = 0 or T(t) = 0
1
That is, we wish to exclude the constant solution u(x, t) = 0.
i ni ti al boundary value problems 309
for all t. The latter possibility is not interesting, because if T(t) = 0 for all t,
then u(x, t) = X(x)T(t) = 0, the trivial solution. Hence, the boundary condition
u(0, t) = 0 implies that X(0) = 0. A similar argument shows that the boundary
condition u(L, t) = 0 forces X(L) = 0.
Step 3: Solve the equation subject to these boundary conditions. We wish to
seek nontrivial solutions of the twopoint boundary value problem
X
//
+ λX = 0, and X(0) = 0 = X(L). (11.23)
The form of the solutions will depend upon the sign of λ, and there are three
possibilities.
Case 1: λ < 0. We claim that if λ < 0, then the boundary value problem (11.23)
has no nontrivial solutions. To see why, suppose that λ = −β
2
where β > 0.
(Introducing β is solely for the purpose of making the solutions look cleaner.)
The ode for X becomes
X
//
− β
2
X = 0.
The associated characteristic equation is m
2
− β
2
= 0, which has distinct, real
roots m = ±β. The general solution of this ode is therefore
X(x) = Ce
−βx
+ De
βx
, (11.24)
where C and D are arbitrary constants. To solve for C and D, we must incorporate
the boundary conditions. Using X(0) = 0 yields 0 = C + D, and using X(L) = 0
yields 0 = Ce
−βL
+ De
βL
. Since D = −C, the latter equation can be written as
C
_
e
−βL
−e
βL
_
= 0.
We may exclude the possibility C = 0, because that would force D = 0, implying
that X(x) = 0 and ultimately leading us to the trivial solution. If C ,= 0, then
e
−βL
−e
βL
= 0. Multiplying both sides by e
βL
and rearranging terms yields
e
2βL
= 1. However, this equality is impossible because 2βL is positive, which
would force e
2βL
> 1.
Case 2: λ = 0. The boundary value problem (11.23) also has no nontrivial
solutions if λ = 0. In this case, the X equation reduces to X
//
= 0. It is easy to
310 separati on of vari ables
solve this ode, integrating twice to obtain the general solution X(x) = Cx + D.
The only way this linear function can possibly satisfy the boundary conditions
X(0) = 0 and X(L) = 0 is if C = D = 0. This means that X(x) = 0 and,
consequently, that u(x, t) = X(x)T(t) = 0 as well. Unfortunately, we have yet to
produce a single interesting solution of our pde and boundary conditions, but
our luck is about to change.
Case 3: λ > 0. Now assume that λ = β
2
where β > 0. (Again, introducing β is
merely for convenience.) The boundary value problem (11.23) becomes
X
//
+ β
2
X = 0, and X(0) = 0 = X(L).
The characteristic equation for the ode is m
2
+ β
2
= 0, which has pure imaginary
roots m = ±βi. Therefore, the general solution of the ode is
X(x) = Ccos(βx) + Dsin(βx),
where C and D are constants. The boundary condition X(0) = 0 implies that
C = 0, and we are left with X(x) = Dsin(βx). The other boundary condition
X(L) = 0 implies that Dsin(βL) = 0. One possibility is D = 0, but this would
lead to the same trivial solution that we are trying to avoid. Fortunately, there
is a much more interesting possibility: sin(βL) = 0 if β is chosen appropriately.
We know that β > 0 by assumption, and L > 0 since it represents the length of
our spatial domain. Therefore, if
βL = nπ (n = 1, 2, 3, . . .),
then we will have found nonzero solutions of the X equation that satisfy both
boundary conditions. In order to index the solutions, let us deﬁne
β
n
=
nπ
L
, λ
n
= β
2
n
(n = 1, 2, 3, . . .)
and
X
n
(x) = D
n
sin(β
n
x) = D
n
sin
_
nπx
L
_
(n = 1, 2, 3 . . .), (11.25)
i ni ti al boundary value problems 311
where D
n
are arbitrary constants. The functions X
n
(x) form the set of all possible
solutions of the boundary value problem (11.23).
Now that we have solved the X equation in (11.22), let us turn our attention to
T
//
+ λc
2
T = 0.
In light of the above calculations, we are only interested in the solution for special
choices of λ, namely for λ
n
= β
2
n
. The equation
T
//
+
n
2
π
2
L
2
c
2
T = 0 (n = 1, 2, 3, . . .)
has characteristic equation
m
2
+
_
nπc
L
_
2
= 0.
The roots
m = ±
_
nπc
L
_
i
are pure imaginary, and therefore the solutions of the T equation have the form
T
n
(t) = E
n
cos
_
nπct
L
_
+ F
n
sin
_
nπct
L
_
(n = 1, 2, 3, . . .). (11.26)
Here, E
n
and F
n
are arbitrary constants.
Step 4: Building the general solution. We will now attempt to build the set
of all possible functions that simultaneously satisfy the pde (11.17) and the
boundary conditions (11.20)–(11.21). The initial conditions will be incorporated
later. Now that we have found all nontrivial solutions X
n
(x) and T
n
(t) of the
odes (11.22) and boundary conditions, recall that u(x, t) = X(x)T(t). Thus, the
functions
u
n
(x, t) = X
n
(x)T
n
(t) (n = 1, 2, 3, . . .) (11.27)
are solutions of the original pde (11.17) which also satisfy both boundary condi
tions (11.20)–(11.21). Since the wave equation (11.17) is linear and homogeneous,
the Superposition Principle 8.1.12 states that any ﬁnite linear combination
N
∑
n=1
G
n
u
n
(x, t)
312 separati on of vari ables
(G
n
are constants) is also a solution of the pde and its boundary conditions.
Surprisingly, we can say much more: the general solution of the pde and its
boundary conditions is given by the inﬁnite sum
∞
∑
n=1
G
n
u
n
(x, t),
provided that this series converges in some sense. Understanding why this inﬁnite
series represents the general solution of the pde with its boundary conditions
requires some effort and is [part of] the subject of the next chapter. For the
moment, we will proceed formally, assuming that this inﬁnite sum really does
represent the solution we seek. Justiﬁcation will be provided later.
Recalling the deﬁnition of u
n
(x, t), we have claimed that the general solution
of the wave equation (11.17) with its two boundary conditions (11.20)–(11.21) is
given by
∞
∑
n=1
G
n
u
n
(x, t) =
∞
∑
n=1
G
n
X
n
(x)T
n
(t)
=
∞
∑
n=1
G
n
_
E
n
cos
_
nπct
L
_
+ F
n
sin
_
nπct
L
__
D
n
sin
_
nπx
L
_
.
Products of arbitrary constants can be combined to make the solution a bit more
concise: let A
n
= G
n
E
n
D
n
and B
n
= G
n
F
n
D
n
to obtain
u(x, t) =
∞
∑
n=1
_
A
n
cos
_
nπct
L
_
+ B
n
sin
_
nπct
L
__
sin
_
nπx
L
_
. (11.28)
Step 5: Use the initial conditions. Armed with the general solution (11.28) of
the wave equation and its two Dirichlet boundary conditions, the ﬁnal step is to
incorporate the initial conditions. Since u(x, 0) = φ(x), setting t = 0 in (11.28)
yields
φ(x) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
. (11.29)
Equation (11.29) may seem troubling and confusing at ﬁrst: it says that our
initial displacement φ must be a sum of sine waves with various frequencies
and amplitudes. However, we never made any assumptions on φ(x) when we
posed our original Dirichlet problem. Would it be a severe restriction to allow
only those initial conditions φ(x) which have special sine series representations
i ni ti al boundary value problems 313
of the form (11.29)? How “big” is the class of functions that have sine series
representations? If a function φ(x) has such a representation, is there a systematic
way to determine constants A
n
? Luckily, the answers to these three questions
are “No”, “Very Big”, and “Yes”, respectively. These issues will be studied in
greater depth during the next chapter.
To use the other initial condition u
t
(x, 0) = ψ(x), we must formally differen
tiate the solution (11.28) with respect to t:
u
t
(x, t) =
∞
∑
n=1
_
−
nπc
L
A
n
sin
_
nπct
L
_
+ B
n
nπc
L
cos
_
nπct
L
__
sin
_
nπx
L
_
.
Now setting t = 0,
ψ(x) = u
t
(x, 0) =
∞
∑
n=1
B
n
nπc
L
sin
_
nπx
L
_
.
As with the other initial condition, we have found that ψ(x) would need to have
a special sine series representation in order for us to declare victory over the
Dirichlet problem (11.17)–(11.21). To summarize: IF there exist constants A
n
and
B
n
such that the initial conditions φ(x), ψ(x) can be represented as convergent
sine series
φ(x) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
and ψ(x) =
∞
∑
n=1
B
n
nπc
L
sin
_
nπx
L
_
on 0 ≤ x ≤ L, then the solution of the Dirichlet problem (11.17)–(11.21) is given
by Equation (11.28) above. Series of this form are called Fourier sine series, and
we will learn more about them in the next chapter.
11.2.2 Heat Equation, Dirichlet Problem. The homogeneous Dirichlet problem
for the heat equation is given by
u
t
= κu
xx
(0 < x < L) (11.30)
u(x, 0) = φ(x) (0 < x < L) (11.31)
u(0, t) = 0 (t ≥ 0) (11.32)
u(L, t) = 0 (t ≥ 0). (11.33)
314 separati on of vari ables
Physically, this models the transfer of heat within a wire of length L, with initial
temperature distribution φ(x) and with both ends of the wire in contact with
zerodegree blocks of ice. We will follow the same steps used to solve the
Dirichlet problem for the wave equation.
Step 1: Separate the variables. Substituting u(x, t) = X(x)T(t) into the pde in
equation (11.30), we have XT
/
= κX
//
T. By algebra,
X
//
X
=
T
/
κT
.
As before, the only way a function of x could be identically equal to a function
of t is if both expressions are equal to a common constant:
X
//
X
=
T
/
κT
= −λ.
Equivalently, we have a system of two odes:
X
//
+ λX = 0 and T
/
+ λκT = 0.
Step 2: Boundary conditions. Using the fact that u(x, t) = X(x)T(t), the homo
geneous Dirichlet boundary conditions (11.32)–(11.33) yield 0 = X(0)T(t) and
0 = X(L)T(t). One possibility is T(t) = 0, but this would lead us to the trivial
solution u(x, t) = 0. Hence, we may focus on the more interesting possibility
that both X(0) = 0 and X(L) = 0.
Step 3: Solve the equation subject to these boundary conditions. Notice that
X satisﬁes the same twopoint boundary value problem (11.23) that we encoun
tered while solving the Dirichlet problem for the wave equation. As before, if
λ ≤ 0 there are no solutions. For λ > 0, it is convenient to let λ = β
2
where
β > 0. The general solution of X
//
+ β
2
X = 0 is X(x) = Ccos(βx) + Dsin(βx),
where C and D are constants. Using the boundary condition X(0) = 0 yields
C = 0, and the boundary condition X(L) = 0 yields Dsin(βL) = 0. We may
assume D ,= 0, because otherwise X(x) is identically zero and we are led to the
trivial solution of the pde. Thus, sin(βL) = 0, and since both β > 0 and L > 0, it
i ni ti al boundary value problems 315
must be the case that βL is a positive integer multiple of π. Deﬁne
β
n
=
nπ
L
λ
n
= β
2
n
=
_
nπ
L
_
2
(n = 1, 2, 3, . . .),
and
X
n
(x) = D
n
sin(β
n
x) = D
n
sin
_
nπx
L
_
(n = 1, 2, 3, . . .).
The functions X
n
(x) satisfy the boundary value problem (11.23). With these
special choices of λ in mind, we return to the T equation
T
/
+ λκT = 0.
Replacing λ with λ
n
,
T
/
+κ
_
nπ
L
_
2
T = 0 (n = 1, 2, 3, . . .),
a ﬁrstorder constantcoefﬁcient ode. By separation of variables, the solutions
are
T
n
(t) = E
n
e
−κ(nπ/L)
2
t
(n = 1, 2, 3, . . .),
where E
n
are constants.
Step 4: Building the general solution. Since u(x, t) = X(x)T(t), let us deﬁne
u
n
(x, t) = X
n
(x)T
n
(t) = A
n
sin
_
nπx
L
_
e
−κ(nπ/L)
2
t
,
where A
n
= D
n
E
n
are constants. Each function u
n
(x, t) satisﬁes the pde (11.30)
and both of its Dirichlet boundary conditions (11.32)–(11.33). The general solu
tion of the pde with its boundary conditions is given by
u(x, t) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
e
−κ(nπ/L)
2
t
. (11.34)
Step 5: Use the initial condition. Finally, it remains to use the initial condi
tion (11.31). Setting t = 0 in our formula for u(x, t), we have
φ(x) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
.
316 separati on of vari ables
Thus, formula (11.34) is the solution of the Dirichlet problem (11.30)–(11.33)
provided that the initial condition φ(x) has a Fourier sine series representation
on the interval 0 < x < L. Given a speciﬁc choice of initial condition φ(x), we
will soon learn how to determine the values of the Fourier coefﬁcients A
n
.
In solving the homogeneous Dirichlet problems for the heat and wave equa
tions, we encountered the twopoint boundary value problem
X
//
= −λX X(0) = 0 = X(L).
Note the similarity between the differential equation
−
d
2
dx
2
X = λX
and the matrix equation Av = λv. The operator −
d
2
dx
2
takes the place of the
matrix A, and the function X(x) takes the place of the vector v. For certain
special choices of λ, namely λ
n
= (nπ/L)
2
, the boundary value problem has
nonzero solutions X. By analogy with linear algebra jargon, the numbers λ
n
are
called eigenvalues and the functions X
n
(x) = sin(nπx/L) are called eigenvectors
or eigenfunctions for the operator −
d
2
dx
2
. The problem −
d
2
X
dx
2
= λX is called an
eigenvalue problem. Notice that in this case, there are inﬁnitely many eigenvalues.
Unlike a square matrix A, the operator −
d
2
dx
2
is an inﬁnitedimensional linear
operator. Much of the branch of mathematics known as functional analysis
involves extending various notions from linear algebra to inﬁnitedimensional
settings.
Example 11.2.1. Suppose r ,= 0 is a constant. Solve
u
t
= κu
xx
−ru (0 < x < L) (11.35)
u(x, 0) = φ(x) (0 < x < L) (11.36)
u(0, t) = 0 (t ≥ 0) (11.37)
u(L, t) = 0 (t ≥ 0). (11.38)
Solution: This is a homogeneous Dirichlet problem. The −ru term in the pde
is known as a reaction term, and the pde itself is an example of a reaction
diffusion equation. We will solve the pde by separation of variables: substituting
u(x, t) = X(x)T(t) into (11.35) yields XT
/
= κX
//
T −rXT. Some algebra will
i ni ti al boundary value problems 317
help us isolate the quantity X
//
/X as in the previous examples. This time, we
ﬁnd that
X
//
X
=
T
/
+ rT
κT
= −λ,
where, as usual, −λ is a constant. Solving the equation X
//
+ λX = 0 subject
to the boundary conditions (11.37)–(11.38) follows precisely the same steps as
before. The only values of λ for which the eigenvalue problem
X
//
+ λX = 0, X(0) = 0 = X(L)
has nonzero solutions are the eigenvalues
λ
n
=
_
nπ
L
_
2
(n = 1, 2, 3, . . .).
The corresponding eigenfunctions are
X
n
(x) = sin
_
nπx
L
_
(n = 1, 2, 3, . . .).
That is, the only nontrivial solutions of the eigenvalue problem are scalar
multiples of these functions X
n
(x).
It remains to solve the T equation T
/
+ rT = −λκT using the special λ values
(the eigenvalues) we found above. The ﬁrstorder constantcoefﬁcient ordinary
differential equation T
/
= (−λκ −r)T has solution
T(t) = Ce
(−λκ−r)t
,
where C is a constant. Using the eigenvalues λ = λ
n
, we are led to deﬁne
T
n
(t) = C
n
e
−rt
e
−λ
n
κt
= C
n
e
−rt
e
−κ(nπ/L)
2
t
(n = 1, 2, 3, . . .).
Finally, we build the general solution by deﬁning u
n
(x, t) = X
n
(x)T
n
(t) and
summing:
u(x, t) =
∞
∑
n=1
B
n
u
n
(x, t),
318 separati on of vari ables
where B
n
are constants. Inserting our expressions for X
n
and T
n
into this
summation and abbreviating A
n
= B
n
C
n
,
u(x, t) = e
−rt
∞
∑
n=1
A
n
sin
_
nπx
L
_
e
−κ(nπ/L)
2
t
(11.39)
is the general solution of the pde (11.35) with boundary conditions (11.37)–
(11.38).
The ﬁnal step is to use the initial condition (11.36) by setting t = 0 in (11.39).
The result is
φ(x) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
,
implying once again that our initial condition φ(x) must have a Fourier sine
series representation.
Take a moment to compare the solutions of the Dirichlet problem for the
heat equation (11.30)–(11.33) and the Dirichlet problem for the reactiondiffusion
system (11.35)–(11.38). The only difference between the pdes (11.30) and (11.35)
is the presence of the reaction term −ru in the latter equation. By comparison,
the only difference between the solutions (11.34) and (11.39) is the presence of
the e
−rt
factor in the latter formula. The solution of u
t
= κu
xx
−ru blends the
solutions of the heat equation u
t
= κu
xx
with the [exponential] solutions of the
ode u
t
= −ru. Roughly speaking, the two terms on the right hand side of (11.35)
seem to act independently of one another, and the overall solution is a hybrid of
the behaviors we would see if either term were absent.
11.2.3 Wave Equation, Neumann Problem. The homogeneous Neumann prob
lem for the wave equation is given by
u
tt
= c
2
u
xx
(0 < x < L) (11.40)
u(x, 0) = φ(x) (0 < x < L) (11.41)
u
t
(x, 0) = ψ(x) (0 < x < L) (11.42)
u
x
(0, t) = 0 (t ≥ 0) (11.43)
u
x
(L, t) = 0 (t ≥ 0). (11.44)
i ni ti al boundary value problems 319
Until we incorporate the boundary conditions, the procedure used to solve this
Neumann problem is identical to how we solved the Dirichlet problem (11.17)–
(11.21).
Step 1: Separate the variables. Seeking separated solutions u(x, t) = X(x)T(t)
of the pde (11.40) leads to
X
//
X
=
T
//
c
2
T
= −λ,
where λ is a constant. The pair of odes
X
//
+ λX = 0 and T
//
+ λc
2
T = 0 (11.45)
is the same pair that we encountered when solving the Dirichlet problem for the
wave equation.
Step 2: Boundary Conditions. Although the ode for X is quite familiar, the
boundary conditions will be different this time. Since u(x, t) = X(x)T(t), note
that u
x
(x, t) = X
/
(x)T(t). Thus, the Neumann conditions (11.43)–(11.44) give
0 = u
x
(0, t) = X
/
(0)T(t) and 0 = u
x
(L, t) = X
/
(L)T(t).
These conditions are certainly satisﬁed if T(t) = 0, but this would result in
the trivial solution u(x, t) = X(x)T(t) = 0. Instead, it must be the case that
X
/
(0) = 0 = X
/
(L).
Step 3: Solve the equation subject to these boundary conditions. We must
seek nontrivial solutions of the eigenvalue problem
X
//
= −λX and X
/
(0) = 0 = X
/
(L). (11.46)
The form of the solutions will depend upon the sign of λ, and there are three
possibilities.
Case 1: λ < 0. We claim that there are no negative eigenvalues—i.e., if λ < 0
then the boundary value problem (11.46) has no nontrivial solutions. Suppose
320 separati on of vari ables
that λ = −β
2
where β > 0. The ode
X
//
− β
2
X = 0.
has associated characteristic equation m
2
− β
2
= 0, which has distinct, real roots
m = ±β. The general solution of this ode is therefore
X(x) = Ce
−βx
+ De
βx
,
where C and D are arbitrary constants. Differentiating this expression yields
X
/
(x) = −βCe
−βx
+ βDe
βx
.
The boundary condition X
/
(0) = 0 tells us that
0 = −βC + βD
and, since β > 0, we conclude that C = D. The other boundary condition states
that
0 = X
/
(L) = −βCe
−βL
+ βDe
βL
.
Since β > 0 and C = D, this equation reduces to Ce
−βL
= Ce
βL
. We may assume
that C ,= 0, because otherwise we would have D = 0 as well, implying that
X(x) = 0 and ultimately leading to the trivial solution u(x, t) = 0. Therefore,
e
−βL
= e
βL
, or equivalently e
2βL
= 1. However, since β and L are positive, it is
impossible for e
2βL
= 1. Thus, there are no negative eigenvalues.
Case 2: λ = 0. Unlike the homogeneous Dirichlet problems we solved in
previous subsections, λ = 0 actually is an eigenvalue—i.e., the boundary value
problem (11.46) does have nontrivial solutions if λ = 0. Integrating the equation
X
//
= 0 twice, the general solution is X(x) = C + Dx and its derivative is
X
/
(x) = D. The boundary conditions X
/
(0) = 0 and X
/
(L) = 0 imply that
D = 0, but there are no restrictions on C. Any constant function X(x) = constant
will satisfy (11.46) if λ = 0. In particular, X
0
(x) = C
0
is an eigenfunction
corresponding to the eigenvalue λ = 0. The corresponding solution T
0
(t) of the
T equation in (11.45) will be obtained later.
i ni ti al boundary value problems 321
Case 3: λ > 0. Now assume that λ = β
2
where β > 0. (Again, introducing β is
merely for convenience.) The boundary value problem (11.23) becomes
X
//
+ β
2
X = 0, and X
/
(0) = 0 = X
/
(L).
The characteristic equation for the ode is m
2
+ β
2
= 0, which has pure imaginary
roots m = ±βi. Therefore, the general solution of the ode is
X(x) = Ccos(βx) + Dsin(βx),
where C and D are constants, and its derivative is
X
/
(x) = −βCsin(βx) + βDcos(βx).
The boundary condition X
/
(0) = 0 implies that D = 0, from which our expres
sions for X and X
/
reduce to
X(x) = Ccos(βx) and X
/
(x) = −βCsin(βx).
The other boundary condition X
/
(L) = 0 implies that −βCsin(βL) = 0. As usual,
we avoid C = 0 and consider the more interesting possibility that sin(βL) = 0.
Since βL > 0, the only way to satisfy this equation is if
βL = nπ (n = 1, 2, 3, . . .).
Deﬁning
β
n
=
nπ
L
and λ
n
= β
2
n
(n = 1, 2, 3, . . .),
we have the same positive eigenvalues λ
n
that we encountered when solving
Dirichlet problems. However, the corresponding eigenfunctions X
n
(x) are cosine
functions instead of sine functions:
X
n
(x) = C
n
cos(β
n
x) = C
n
cos
_
nπx
L
_
(n = 1, 2, 3 . . .), (11.47)
where C
n
are arbitrary constants. The functions X
n
(x) form the set of all possible
solutions of the boundary value problem (11.46).
322 separati on of vari ables
With these positive eigenvalues in mind, consider the T equation in (11.45).
The equation
T
//
+
n
2
π
2
L
2
c
2
T = 0 (n = 1, 2, 3, . . .)
has characteristic equation
m
2
+
_
nπc
L
_
2
= 0.
The roots
m = ±
_
nπc
L
_
i
are pure imaginary, and therefore the solutions of the T equation have the form
T
n
(t) = E
n
cos
_
nπct
L
_
+ F
n
sin
_
nπct
L
_
(n = 1, 2, 3, . . .), (11.48)
where E
n
and F
n
are arbitrary constants.
Now, recall that λ = 0 is also an eigenvalue. In that case, the T equation
reduces to T
//
= 0, and this ode can be solved by integrating twice with respect
to t. The solution has the form
T
0
(t) = E
0
+ F
0
t,
where E
0
and F
0
are arbitrary constants.
Step 4: Building the general solution. The general solution of the wave equa
tion (11.40) with the two Neumann boundary conditions (11.43)–(11.44) is con
structed in the usual way, by taking an inﬁnite linear combination of the separated
solutions that we found above. Deﬁne
u
n
(x, t) = X
n
(x)T
n
(t) (n = 0, 1, 2, . . .),
noting that n = 0 must be included this time because λ = 0 is an eigenvalue.
The general solution is
u(x, t) =
∞
∑
n=0
G
n
u
n
(x, t),
where G
n
are constants. Because the form of u
0
(x, t) is different than the other
terms in this summation, it is useful to present the general solution in the
i ni ti al boundary value problems 323
following way:
u(x, t) = G
0
u
0
(x, t) +
∞
∑
n=1
G
n
u
n
(x, t) = G
0
C
0
(E
0
+ F
0
t)
+
∞
∑
n=1
G
n
_
E
n
cos
_
nπct
L
_
+ F
n
sin
_
nπct
L
__
C
n
cos
_
nπx
L
_
.
This expression simpliﬁes if we abbreviate various combinations of constants. For
n ≥ 1, let A
n
= G
n
E
n
C
n
and B
n
= G
n
F
n
C
n
. For n = 0, we do something slightly
different: let G
0
C
0
E
0
=
1
2
A
0
and G
0
C
0
F
0
=
1
2
B
0
. The reason for including the
factors of
1
2
will be explained below. Overall, we have shown that the general
solution of the pde and the two Neumann conditions is
u(x, t) =
A
0
2
+
B
0
2
t
+
∞
∑
n=1
_
A
n
cos
_
nπct
L
_
+ B
n
sin
_
nπct
L
__
cos
_
nπx
L
_
.
(11.49)
Step 5: Use the initial conditions. Since u(x, 0) = φ(x), setting t = 0 in (11.49)
implies that
φ(x) =
A
0
2
+
∞
∑
n=1
A
n
cos
_
nπx
L
_
. (11.50)
Equation (11.50) is in the standard form of a Fourier cosine series. The factor of
1
2
in front of the leading term will be explained when we study Fourier series
in the next chapter. As we shall see, insisting that φ(x) have a Fourier cosine
series representation is not a severe restriction at all. Moreover, we will soon
develop a systematic procedure for determining the Fourier coefﬁcients A
n
for a
function φ(x) deﬁned on an interval 0 ≤ x ≤ L. To use the other initial condition
u
t
(x, 0) = ψ(x), we must formally
2
differentiate (11.49) with respect to t:
u
t
(x, t) =
B
0
2
+
∞
∑
n=1
nπc
L
_
−A
n
sin
_
nπct
L
_
+ B
n
cos
_
nπct
L
__
cos
_
nπx
L
_
.
Setting t = 0, we have
ψ(x) = u
t
(x, 0) =
B
0
2
+
∞
∑
n=1
nπc
L
B
n
cos
_
nπx
L
_
. (11.51)
2
Strictly speaking, termbyterm differentiation of the series (11.49) is justiﬁed only under certain
minor assumptions regarding the convergence of the series.
324 separati on of vari ables
Thus, we must require that the initial velocity ψ(x) also have a Fourier co
sine series representation. In summary, the solution of the homogeneous Neu
mann problem (11.40)–(11.44) is given by formula (11.49), provided that the
initial conditions φ(x) and ψ(x) have Fourier cosine series representations (11.50)
and (11.51).
11.2.4 Heat Equation, Neumann Problem. The homogeneous Neumann prob
lem for the heat equation is given by
u
t
= κu
xx
(0 < x < L) (11.52)
u(x, 0) = φ(x) (0 < x < L) (11.53)
u
x
(0, t) = 0 (t ≥ 0) (11.54)
u
x
(L, t) = 0 (t ≥ 0). (11.55)
Physically, this problem models the diffusion of heat within a onedimensional
wire whose ends are insulated to prevent heat ﬂux across the boundaries. The
general solution of the pde with its two Neumann conditions can be presented as
u(x, t) =
A
0
2
+
∞
∑
n=1
A
n
e
−(nπ/L)
2
κt
cos
_
nπx
L
_
. (11.56)
The reader is encouraged to use the separation of variables technique to derive
formula (11.56). As in the Neumann problem for the wave equation, note that the
initial condition φ(x) is required to have a Fourier cosine series representation of
the form (11.50). For any reasonably wellbehaved choice of initial condition φ(x)
(e.g., if φ is continuous on 0 ≤ x ≤ L), then there is a straightforward procedure
for calculating the Fourier coefﬁcients A
n
(see next chapter).
11.2.5 Mixed Boundary Conditions: An Example. As a ﬁnal illustration of the
separation of variables technique, we will solve the heat equation with mixed
boundary conditions:
u
t
= κu
xx
(0 < x < L) (11.57)
u(x, 0) = φ(x) (0 < x < L) (11.58)
u(0, t) = 0 (t ≥ 0) (11.59)
u
x
(L, t) = 0 (t ≥ 0). (11.60)
i ni ti al boundary value problems 325
Physically, the Dirichlet condition at x = 0 simulates placing the x = 0 end of
the wire in contact with a block of ice, holding the temperature constant at zero
degrees Celsius. The Neumann condition at x = L simulates insulating that end
of the wire, thereby preventing heat from entering/exitting the wire. If we seek
separated solutions u(x, t) = X(x)T(t), we ﬁnd that X and T must satisfy odes
of the form
X
//
+ λX = 0 and T
/
+κλT = 0,
where λ is a constant. The boundary conditions (11.59)–(11.60) imply that
X(0) = 0 and X
/
(L) = 0. The reader should show that there are no negative
eigenvalues—that is, if λ < 0, then it is impossible to simultaneously satisfy
the ode for X as well as both of these boundary conditions. We also claim that
λ = 0 is not an eigenvalue. If λ = 0, the ode for X reduces to X
//
= 0 which has
general solution X(x) = Cx + D, where C and D are constants. The condition
X(0) = 0 implies that D = 0, and the condition X
/
(L) = 0 implies that C = 0
as well. Consequently, X(x) = 0 which leads to u(x, t) = 0, the trivial solution.
Finally, let us seek positive eigenvalues by setting λ = β
2
where β > 0. The
general solution of the X equation is
X(x) = Ccos(βx) + Dsin(βx).
From the boundary condition X(0) = 0, we conclude that C = 0 and the
expression for X(x) reduces to X(x) = Dsin(βx). To use the other boundary
condition, we ﬁrst compute the derivative X
/
(x) = βDcos(βx). Since X
/
(L) = 0,
we obtain 0 = βDcos(βL). We know that β > 0 by assumption and, as usual, we
can rule out D = 0. However, if β is chosen such that cos(βL) = 0, then we will
have produced nontrivial solutions of the boundary value problem for X. The
only possible choices are
βL = −
π
2
+ nπ (n = 1, 2, 3, . . .).
Deﬁning
β
n
=
_
n −
1
2
_
π
L
(n = 1, 2, 3, . . .),
326 separati on of vari ables
the eigenvalues are given by λ
n
= β
2
n
. The corresponding solutions of the X
equation are
X
n
(x) = D
n
sin
_
_
_
n −
1
2
_
πx
L
_
_
(n = 1, 2, 3, . . .),
where D
n
are constants.
We now turn our attention to the T equation. The general solution of the
ﬁrstorder ordinary differential equation T
/
+κλT = 0 is given by
T(t) = Be
−κλt
,
where B is a constant. The only λ values of interest are the eigenvalues λ
n
, which
motivates us to deﬁne
T
n
(t) = B
n
e
−κλ
n
t
= B
n
e
−κ[(n−
1
2
)π/L]
2
t
.
Letting u
n
(x, t) = X
n
(x)T
n
(t) for n ≥ 1, the general solution of heat equation
with our mixed boundary conditions is
u(x, t) =
∞
∑
n=1
F
n
u
n
(x, t) =
∞
∑
n=1
F
n
B
n
e
−κ[(n−
1
2
)π/L]
2
t
D
n
sin
_
_
_
n −
1
2
_
πx
L
_
_
.
Combining the various constants by introducing A
n
= F
n
B
n
D
n
, we have
u(x, t) =
∞
∑
n=1
A
n
e
−κ[(n−
1
2
)π/L]
2
t
sin
_
_
_
n −
1
2
_
πx
L
_
_
.
Finally, the initial condition (11.58) implies that
φ(x) =
∞
∑
n=1
A
n
sin
_
_
_
n −
1
2
_
πx
L
_
_
.
This expression for φ(x) is not quite in the form of a Fourier sine series. As we
shall see, Fourier sine series and Fourier cosine series are special cases of a more
i ni ti al boundary value problems 327
general class of series: (full) Fourier series. For some functions φ(x), we will want
to use (full) Fourier series representations instead of Fourier sine or cosine series.
Exercises
1. Solve the homogeneous Neumann problem for the heat equation on a halfline:
u
t
= κu
xx
(0 < x < ∞)
u(x, 0) = φ(x) (0 < x < ∞)
u
x
(0, t) = 0.
2. Solve the homogeneous Neumann problem for the wave equation on a half
line:
u
tt
−c
2
u
xx
= 0 (0 < x < ∞)
u(x, 0) = φ(x) (0 < x < ∞)
u
t
(x, 0) = ψ(x) (0 < x < ∞)
u
x
(0, t) = 0.
In case it helps reduce the number of cases you must consider, just give the
solution for t ≥ 0.
3. Solve the heat equation on a halfline with an inhomogeneous Dirichlet
condition:
u
t
= κu
xx
(0 < x < ∞)
u(x, 0) = φ(x) (0 < x < ∞)
u(0, t) = g(t).
To do so, ﬁrst let w(x, t) = u(x, t) −g(t) and show that w satisﬁes an inhomo
geneous pde with a homogeneous Dirichlet boundary condition:
w
t
= κw
xx
+ f (x, t) (0 < x < ∞)
w(x, 0) =
˜
φ(x) (0 < x < ∞)
w(0, t) = 0,
328 separati on of vari ables
where
˜
φ(x) = φ(x) −g(0) and f (x, t) = −g
/
(t). Then, solve for w by combin
ing the oddreﬂection method with Duhamel’s Principle (see (10.31)). Finally,
obtain the overall solution by recalling that u(x, t) = w(x, t) + g(t).
4. Solve the heat equation on a halfline with an inhomogeneous Neumann
condition:
u
t
= κu
xx
(0 < x < ∞)
u(x, 0) = φ(x) (0 < x < ∞)
u
x
(0, t) = g(t).
To do so, ﬁrst let w(x, t) = u(x, t) −xg(t) and mimic the procedure outlined
in the previous exercise.
5. Unfortunately, the separation of variables method does not work for all linear,
constantcoefﬁcient pdes. For example, consider the homogeneous Dirichlet
problem for the transport equation:
u
t
−cu
x
= 0 (0 < x < L),
u(0, t) = 0
u(L, t) = 0,
where c is a positive constant. Show that there are no (nonzero) separated
solutions of this problem. That is, there are no eigenvalues.
6. Use separation of variables to solve the homogeneous Neumann problem for
the heat equation:
u
t
= κu
xx
(0 < x < L)
u(x, 0) = φ(x) (0 < x < L)
u
x
(0, t) = 0
u
x
(L, t) = 0.
Your answer should be expressed in the form of a cosine series.
7. The wave equation can be used to model a vibrating string in a “vacuum” (i.e.,
we neglect friction/air resistance). The following Dirichlet problem models a
i ni ti al boundary value problems 329
vibrating string in a resistant medium:
u
tt
= c
2
u
xx
−ru
t
(0 < x < L)
u(x, 0) = φ(x) (0 < x < L)
u
t
(x, 0) = ψ(x) (0 < x < L)
u(0, t) = 0
u(L, t) = 0.
Assuming that r is a constant and 0 < r < 2πc/L, use separation of variables
to ﬁnd a series solution to this Dirichlet problem.
8. The purpose of this problem is to solve the heat equation on a onedimensional
ringshaped domain of total length 2L. The idea is to solve the heat equation
on the domain −L ≤ x ≤ L, where x = −L and x = L correspond to the same
physical location. To simulate this, we set up periodic boundary conditions as
follows:
u
t
= κu
xx
(−L ≤ x ≤ L)
u(−L, t) = u(L, t)
u
x
(−L, t) = u
x
(L, t).
First, show that the eigenvalues are λ
n
= (nπ/L)
2
for n ≥ 0. (Note that 0 is
an eigenvalue in this case.) Then, show that the solution of this problem can
be written in the form
u(x, t) =
A
0
2
+
∞
∑
n=1
_
A
n
cos
_
nπx
L
_
+ B
n
sin
_
nπx
L
__
e
−(nπ/L)
2
κt
.
9. Solve the wave equation with periodic boundary conditions:
u
tt
= c
2
u
xx
(−L < x < L)
u(−L, t) = u(L, t)
u
x
(−L, t) = u
x
(L, t),
where c ,= 0 is a constant.
CHAPTER 12
Introduction to Fourier Series
I
n the previous chapter, we developed the separation of variables technique
to construct the solutions of homogeneous Dirichlet and Neumann problems.
In each example, we were able to construct series representations of the solutions
provided that the initial conditions themselves had special series representations
(i.e., Fourier sine and cosine series). In this chapter, we will study Fourier series
in greater depth, addressing three principal questions:
Which functions φ(x) have Fourier series representations?
Given a function φ(x) that does have such a representation, how can we
calculate the coefﬁcients that appear within the series?
Can we really be sure that these series converge to φ(x)?
Before beginning our study of Fourier series, let us brieﬂy recall another
type of series representation that you likely studied in calculus: Taylor series.
Suppose that a function f (x) has inﬁnitely many derivatives in some open
interval containing the point x = a. In calculus, you learned that the Taylor series
for f (x) centered at a is given by
∞
∑
n=0
f
(n)
(a)
n!
(x −a)
n
.
Computing the constants f
(n)
(a) requires that we ﬁrst calculate the nth derivative
of f and then evaluate at x = a. For example, suppose we wish to compute the
Taylor series for f (x) = e
x
centered at 0. Since f
(n)
(x) = e
x
for all n, it follows
that f
(n)
(0) = 1 for all n. Thus, the Taylor series for e
x
centered at 0 is given by
∞
∑
n=0
1
n!
x
n
.
330
i ntroducti on to fouri er seri es 331
The ratio test shows that this series converges for all real x. Moreover, the series
really does converge to e
x
for all real x.
Recall that an inﬁnite series is deﬁned as a limit of partial sums; e.g.,
e
x
= lim
N→∞
N
∑
n=1
1
n!
x
n
.
Notice that partial sums of a Taylor series are nothing more than polynomials,
as are the factors (x −a)
n
that appear in each individual term. In this sense, a
Taylor series essentially represents a function f (x) as a sum of polynomials.
Fourier series offer another way of representing functions as inﬁnite series. Un
like Taylor series, which use polynomials as “building blocks”, Fourier series are
sums of sine and cosine functions. More speciﬁcally, a Fourier series effectively
decomposes a function φ(x) into a sum of sine and cosine functions all of which
have frequencies that are integer multiples of some “fundamental frequency”.
Deﬁnition 12.0.2. Let L > 0. A Fourier sine series is a series of the form
∞
∑
n=1
A
n
sin
_
nπx
L
_
(0 < x < L). (12.1)
A Fourier cosine series is a series of the form
A
0
2
+
∞
∑
n=1
A
n
cos
_
nπx
L
_
(0 < x < L). (12.2)
A (full) Fourier series is a series of the form
A
0
2
+
∞
∑
n=1
_
A
n
cos
_
nπx
L
_
+ B
n
sin
_
nπx
L
__
(−L < x < L). (12.3)
Notice that the interval over which the full Fourier series is deﬁned is symmetric
about x = 0, whereas the sine and cosine series are deﬁned for (0 < x < L).
Before we tackle the theoretical questions regarding which functions φ(x) have
convergent Fourier series representations, we will explain how to calculate the
coefﬁcients assuming we are given an appropriate φ(x).
332 fouri er seri es
12.1. Fourier series
Computing the coefﬁcients in a Taylor series for φ(x) requires that we calculate
all of the derivatives of φ(x). By contrast, computing Fourier coefﬁcients will
require that we calculate integrals instead.
12.1.1 Fourier sine series. Suppose that φ(x) is deﬁned and integrable on the
interval 0 ≤ x ≤ L, and assume that φ(x) has a convergent Fourier sine series
representation
φ(x) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
(0 < x < L). (12.4)
Our goal is to ﬁnd the values of the Fourier coefﬁcients A
n
. First, we need a
Lemma:
Lemma 12.1.1. Suppose m and n are positive integers. Then
_
L
0
sin
_
mπx
L
_
sin
_
nπx
L
_
dx =
_
_
_
0 if m ,= n
L/2 if m = n.
Proof. The idea is to use some trigonometric identities to simplify the integrand.
First, suppose m = n. The integral becomes
_
L
0
sin
2
_
nπx
L
_
dx,
and the identity sin
2
θ =
1
2
[1 −cos(2θ)] leads us to
1
2
_
L
0
1 −cos
_
2nπx
L
_
dx =
1
2
_
x −
L
2nπ
sin
_
2nπx
L
__¸
¸
¸
¸
L
0
=
L
2
.
The case m ,= n requires a more creative use of trigonometric identities. Speciﬁ
cally, we will use
sin α sin β =
1
2
cos(α − β) −
1
2
cos(α + β). (12.5)
i ntroducti on to fouri er seri es 333
As an exercise, you may wish to verify (12.5) using the more familiar “double
angle identity”
cos(α + β) = cos α cos β −sin α sin β.
Using (12.5) with α = mπx/L and β = nπx/L, we can rewrite our integral as
_
L
0
sin
_
mπx
L
_
sin
_
nπx
L
_
dx =
1
2
_
L
0
cos
_
(m−n)πx
L
_
−cos
_
(m + n)πx
L
_
dx.
With the integrand expressed in this more convenient form, we evaluate
1
2
L
(m−n)π
sin
_
(m−n)πx
L
_
−
1
2
L
(m + n)π
sin
_
(m + n)πx
L
_¸
¸
¸
¸
L
0
=
1
2
L
(m−n)π
sin [(m−n)π] −
1
2
L
(m + n)π
sin [(m + n)π] = 0,
where we have used the fact that the sine of an integer multiple of π is always
zero. The assumptions that m and n are positive integers and m ,= n were
important when writing down this antiderivative. Otherwise, we could not
guarantee that both m−n ,= 0 and m + n ,= 0.
Armed with Lemma 12.1.1, we can calculate the Fourier coefﬁcients in the sine
series representation of φ(x). Choose any positive integer m ≥ 1 and multiply
both sides (12.4) by sin(mπx/L):
φ(x) sin
_
mπx
L
_
=
∞
∑
n=1
A
n
sin
_
nπx
L
_
sin
_
mπx
L
_
.
Now integrate both sides over the interval 0 ≤ x ≤ L:
_
L
0
φ(x) sin
_
mπx
L
_
dx =
_
L
0
∞
∑
n=1
A
n
sin
_
nπx
L
_
sin
_
mπx
L
_
dx
=
∞
∑
n=1
A
n
_
L
0
sin
_
nπx
L
_
sin
_
mπx
L
_
dx.
(When interchanging the sum and the integral, we have tacitly assumed that
the Fourier sine series for φ(x) converges in an appropriate manner.) By
Lemma 12.1.1, all of the integrals in the summand are zero except in the case the
334 fouri er seri es
n = m. Only one term in the summation survives:
_
L
0
φ(x) sin
_
mπx
L
_
dx =
L
2
A
m
.
Therefore, the Fourier sine series coefﬁcients are given by
A
m
=
2
L
_
L
0
φ(x) sin
_
mπx
L
_
dx (m ≥ 1). (12.6)
Example 12.1.2. Calculate the Fourier sine series for φ(x) = sin(2πx) on the
interval 0 < x < 1. Solution: In this case, the length of the interval is L = 1, so
the Fourier sine series should have the form
φ(x) =
∞
∑
n=1
A
n
sin(nπx).
The coefﬁcients A
n
are given by (12.6):
A
n
= 2
_
1
0
φ(x) sin (nπx) dx = 2
_
1
0
sin (nπx) sin (2πx) dx.
Applying Lemma 12.1.1 with L = 1, we have
2
_
1
0
sin (nπx) sin (2πx) dx =
_
_
_
1 if n = 2
0 if n ,= 2.
It follows that A
2
= 1 and A
n
= 0 if n ,= 2. This is not at all surprising, as the
function φ(x) = sin(2πx) is already in the form of a Fourier sine series.
Example 12.1.3. Find the Fourier sine series for the constant function φ(x) = 1
on the interval 0 < x < 1. Solution: Again, L = 1 and the Fourier sine series has
the form
φ(x) =
∞
∑
n=1
A
n
sin(nπx).
The coefﬁcients A
n
are given by (12.6):
A
n
= 2
_
1
0
φ(x) sin (nπx) dx = 2
_
1
0
sin (nπx) dx = −
2
nπ
cos (nπx)
¸
¸
¸
¸
1
0
= −
2
nπ
[cos(nπ) −1] .
i ntroducti on to fouri er seri es 335
Notice that cos(nπ) alternates:
cos(nπ) = (−1)
n
=
_
_
_
1 if n is even
−1 if n is odd.
Therefore, the Fourier sine series coefﬁcients are
A
n
= −
2
nπ
[(−1)
n
−1] =
2
nπ
[1 −(−1)
n
] =
_
_
_
4/nπ if n is odd
0 if n is even,
and the Fourier series for φ(x) is
1 = φ(x) =
4
π
_
sin(πx) +
1
3
sin(3πx) +
1
5
sin(5πx) + · · ·
_
=
4
π
∞
∑
n=1
1
2n −1
sin[(2n −1)πx]. (12.7)
To visualize the convergence of the series (12.7), it is useful to plot the ﬁrst few
partial sums. The left panel of Figure 12.1 shows the ﬁrst three partial sums, and
the right panel shows the 20th partial sum. Notice that the sequence of partial
sums does appear to converge to the function φ(x) = 1 everywhere except at the
endpoints. In fact, when x = 0 or x = 1, all of the terms in the sine series are
equal to zero. Although the Fourier sine series cannot converge to φ(x) at the
endpoints of the interval, we will soon see that this is not an issue for the Fourier
cosine series of φ(x) = 1.
As a side note, formula (12.7) actually provides some rather curious identities.
For example, if we set x = 1/2, then the series becomes
1 =
4
π
_
1 −
1
3
+
1
5
−
1
7
+
1
9
+ . . .
_
.
Equivalently,
π = 4
∞
∑
n=0
(−1)
n
1
2n +1
,
which is a series representation of an important mathematical constant.
336 fouri er seri es
0
1
0
1
x
0
1
0
1
x
4/π 4/π
Figure 12.1. Left panel: The ﬁrst, second, and third partial sums of
the series (12.7). Right panel: The sum of the ﬁrst 20 terms of the
series (12.7). Horizontal lines corresponding to values of 1 and 4/π are
included for reference.
Example 12.1.4. In the previous chapter, we showed that the solution of the
homogeneous Dirichlet problem for the heat equation (11.30)–(11.33) is given by
u(x, t) =
∞
∑
n=1
A
n
e
−κ(nπ/L)
2
t
sin
_
nπx
L
_
.
Setting t = 0, the initial condition (11.31) tells us that
φ(x) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
.
As this is a Fourier sine series, the coefﬁcients are given by formula (12.6).
Assuming that φ(x) has such a representation, we are able to express the solution
of this Dirichlet problem in terms of φ itself:
u(x, t) =
2
L
∞
∑
n=1
e
−κ(nπ/L)
2
t
sin
_
nπx
L
_
_
L
0
φ(x) sin
_
nπx
L
_
dx.
Let us consider a speciﬁc initial condition φ(x) = sin(2πx) with L = 1 (that is,
the spatial domain is 0 ≤ x ≤ 1). In an example above, we showed that the
Fourier coefﬁcients for this function are A
2
= 1 and A
n
= 0 for all n ,= 2. Only
one of the terms in the formula for u(x, t) is nonzero and, for this speciﬁc choice
of initial condition, we have
u(x, t) = e
−κ(2π)
2
t
sin(2πx).
i ntroducti on to fouri er seri es 337
Visualizing the behavior of this solution is straightforward since the time
dependent and spacedependent parts are separated. As t increases, the ex
ponential factor decays to 0, effectively reducing the amplitude of the sinusoidal
heat distribution. More speciﬁcally,
lim
t→∞
u(x, t) = 0
for each x in the domain 0 ≤ x ≤ 1. As an exercise, you may wish to choose some
[positive] value for the diffusion constant κ and use a calculator or computer to
plot u(x, t).
Example 12.1.5. In the previous chapter, we showed that the general solution of
the wave equation (11.17) with homogeneous Dirichlet conditions (11.20)–(11.21)
is given by
u(x, t) =
∞
∑
n=1
_
A
n
cos
_
nπct
L
_
+ B
n
sin
_
nπct
L
__
sin
_
nπx
L
_
.
Using the initial conditions (11.18)–(11.19), we found that
φ(x) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
and ψ(x) =
∞
∑
n=1
nπc
L
B
n
sin
_
nπx
L
_
.
Since φ(x) is in the form of a Fourier sine series, we know that the coefﬁcients
A
n
are
A
n
=
2
L
_
L
0
φ(x) sin
_
nπx
L
_
dx (n = 1, 2, 3 . . .).
If we introduce
˜
B
n
= (nπc/L)B
n
, then the series for ψ(x) also takes the standard
form of a Fourier sine series. We conclude that
˜
B
n
=
2
L
_
L
0
ψ(x) sin
_
nπx
L
_
dx (n = 1, 2, 3 . . .),
from which it follows that
B
n
=
2
nπc
_
L
0
ψ(x) sin
_
nπx
L
_
dx (n = 1, 2, 3 . . .).
Now that we have expressed the coefﬁcients A
n
and B
n
in terms of the given
initial data, we have completely solved the Dirichlet problem (11.17)–(11.21).
338 fouri er seri es
12.1.2 Fourier cosine series. When we solved the homogeneous Neumann prob
lems for the heat and wave equations, we found that the initial conditions
must have convergent Fourier cosine series representations. The process of
determining the Fourier coefﬁcients in the expansion
φ(x) =
A
0
2
+
∞
∑
n=1
A
n
cos
_
nπx
L
_
(12.8)
is very similar to what we did for sine series expansions. First, a technical
Lemma:
Lemma 12.1.6. Suppose m and n are nonnegative integers and L > 0. Then
_
L
0
cos
_
mπx
L
_
cos
_
nπx
L
_
dx =
_
¸
¸
¸
_
¸
¸
¸
_
0 if m ,= n
L/2 if m = n ,= 0
L if m = n = 0.
Proof. The proof is essentially the same as that of Lemma 12.1.1. To handle the
ﬁrst case, use the trigonometric identity
cos α cos β =
1
2
cos(α − β) +
1
2
cos(α + β).
For the second case, use the identity
cos
2
α =
1
2
[1 +cos(2α)] ,
which is actually a special case of the other identity in which α = β. For the ﬁnal
case, notice that the integrand reduces to 1 if m = n = 0.
To determine the Fourier cosine series coefﬁcients in (12.8), we use a similar
trick as with Fourier sine series. Choose any nonnegative integer m and multiply
both sides of (12.8) by cos(mπx/L) to obtain
φ(x) cos
_
mπx
L
_
=
A
0
2
cos
_
mπx
L
_
+
∞
∑
n=1
A
n
cos
_
nπx
L
_
cos
_
mπx
L
_
.
Now integrate both sides over the interval 0 ≤ x ≤ L:
i ntroducti on to fouri er seri es 339
_
L
0
φ(x) cos
_
mπx
L
_
dx =
_
L
0
A
0
2
cos
_
mπx
L
_
dx
+
∞
∑
n=1
A
n
_
L
0
cos
_
nπx
L
_
cos
_
mπx
L
_
dx.
First, suppose m > 0. Then according to Lemma (12.1.6), the only nonzero term
in this series occurs when n = m. Thus, the equation reduces to
_
L
0
φ(x) cos
_
mπx
L
_
dx =
L
2
A
m
if m > 0. If m = 0, then
_
L
0
φ(x) cos
_
mπx
L
_
dx =
_
L
0
A
0
2
cos
_
0πx
L
_
dx =
_
L
0
A
0
2
dx =
L
2
A
0
.
Multiplying both sides of these equations by 2/L, we have shown that the Fourier
cosine series coefﬁcients are given by
A
n
=
2
L
_
L
0
φ(x) cos
_
nπx
L
_
dx (n = 0, 1, 2, . . .). (12.9)
The reason for including the factor of 1/2 in front of the coefﬁcient A
0
in
the Fourier cosine series is now evident. Namely, because the integral in
Lemma 12.1.6 is twice as large if m = n = 0 than if m = n ,= 0, including
the 1/2 in front of A
0
lets us avoid listing multiple cases in formula (12.9).
Example 12.1.7. Find the Fourier cosine series representation for φ(x) = 1 on
the interval 0 ≤ x ≤ 1. Solution: Since L = 1, the Fourier cosine coefﬁcients are
given by
A
n
= 2
_
L
0
cos (nπx) dx =
_
_
_
2 if n = 0
0 if n ,= 0.
In other words, the only nonzero term in the Fourier cosine series for φ(x) = 1
is the leading term: φ(x) = A
0
/2 = 1. This makes sense, because φ(x) is already
in the form of a cosine series: φ(x) = 1 = cos(0πx).
Compare this example with the earlier example in which we calculated the
Fourier sine series representation for φ(x) = 1. For this particular choice of φ(x),
why is the Fourier cosine series so much cleaner than the sine series? We will
explore this question in a later section.
340 fouri er seri es
Example 12.1.8. Find the Fourier cosine series for φ(x) = x on the interval
0 < x < 2. Solution: Using formula (12.9) with L = 2,
A
n
=
_
2
0
x cos
_
nπx
2
_
dx.
We need to handle the n = 0 case separately, because the cosine function in the
integrand reduces to 1 (which affects the form of our antiderivative):
A
0
=
_
2
0
x dx =
x
2
2
¸
¸
¸
¸
2
0
= 2.
For n ≥ 1, we integrate by parts:
A
n
=
2
nπ
x sin
_
nπx
2
_
¸
¸
¸
¸
2
0
−
2
nπ
_
2
0
sin
_
nπx
2
_
dx
=
2
nπ
x sin
_
nπx
2
_
¸
¸
¸
¸
2
0
+
_
2
nπ
_
2
cos
_
nπx
2
_
¸
¸
¸
¸
¸
2
0
.
Since sin(nπ) = 0 and cos(nπ) = (−1)
n
for all integers n, substituting in the
limits of integration yields
A
n
=
_
_
_
0 if n ≥ 1 is even
−8/(nπ)
2
if n ≥ 1 is odd.
The Fourier cosine series for φ(x) = x on the interval 0 < x < 2 is
φ(x) =
A
0
2
+
∞
∑
n=1
A
n
cos
_
nπx
2
_
= 1 −
8
π
2
cos
_
πx
2
_
−
8
(3π)
2
cos
_
3πx
2
_
−
8
(5π)
2
cos
_
5πx
2
_
−· · ·
= 1 −8
∞
∑
n=1
1
[(2n −1)π]
2
cos
_
(2n −1)πx
2
_
.
You may wish to use a computer or calculator to plot the ﬁrst few partial sums
of this cosine series in order to visualize the convergence.
i ntroducti on to fouri er seri es 341
Example 12.1.9. Solve the Neumann problem
u
t
= 2κu
xx
(0 < x < 7)
u(x, 0) = 10 +3 cos
_
πx
7
_
(0 < x < 7)
u
x
(0, t) = 0 (t ≥ 0)
u
x
(L, t) = 0 (t ≥ 0).
Solution: This Neumann problem models heat transfer in a onedimensional
wire of length L = 7 in which the ends are insulated. The diffusion coefﬁcient is
κ = 2. The initial condition is illustrated in Figure 12.2—notice that initially the
highest temperature is 13 (when x = 0) and the lowest temperature is 7 (when
x = 7). According to formula (11.56), the general solution of the pde together
with these two Neumann boundary conditions is
u(x, t) =
A
0
2
+
∞
∑
n=1
A
n
e
−2(nπ/7)
2
t
cos
_
nπx
7
_
.
To incorporate the initial condition, we set t = 0 to ﬁnd that
φ(x) = 10 +3 cos
_
πx
7
_
=
A
0
2
+
∞
∑
n=1
A
n
cos
_
nπx
7
_
,
a Fourier cosine series. Luckily, our initial condition is already in the form of a
Fourier cosine series with A
0
= 20, A
1
= 3, and A
n
= 0 for n ≥ 2. Inserting
these coefﬁcients into the formula for u(x, t), we ﬁnd that the solution of our
Neumann problem is given by
u(x, t) = 10 + 3e
−2(π/7)
2
t
cos
_
πx
7
_
.
Observe that for all x in the domain 0 ≤ x ≤ 7, the temperature approaches 10
as t increases. More precisely,
lim
t→∞
u(x, t) = 10
for all x ∈ [0, 7], because the exponential factor in u(x, t) decays to zero while
the cosine factor remains bounded between −1 and 1. Physically, the heat
equilibrates within the wire, with the temperature converging to the average of
342 fouri er seri es
0
10
13
7
u
(
x
,
t
)
x
0 7
steady−state
temperature
profile
initial temperature
profile (t = 0)
Figure 12.2. Initial (t = 0) and steadystate (t = ∞) heat distributions
for the homogeneous Neumann problem example (see text). Because
the boundaries are insulated, the temperature distribution eventually
equilibrates to the average value of the initial temperature distribution.
the initial temperature proﬁle (see Figure 12.2). This makes sense given that the
ends of the wire are insulated to prevent heat from entering or exiting the wire
at the boundaries.
Even if we had failed to notice that the initial condition φ(x) had the form of a
Fourier cosine series, we could easily calculate the Fourier cosine coefﬁcients by
evaluating the integral (12.9) with the help of Lemma 12.1.6.
12.1.3 Fourier series. Recall that the Fourier series expansion for a function φ(x)
on an interval −L < x < L takes the form
φ(x) =
A
0
2
+
∞
∑
n=1
_
A
n
cos
_
nπx
L
_
+ B
n
sin
_
nπx
L
__
.
Again, notice that unlike the sine and cosine series we considered above, the
Fourier series expansion is deﬁned on an interval that is symmetric about x = 0.
Calculating the Fourier coefﬁcients A
n
and B
n
is accomplished via the same
sort of procedure that we used to compute the coefﬁcients in sine and cosine
series. Suppose that m and n are nonnegative integers and that L > 0. As an
exercise, you can show that, for all m and n,
_
L
−L
sin
_
mπx
L
_
cos
_
nπx
L
_
dx = 0,
i ntroducti on to fouri er seri es 343
if m ,= n, then
_
L
−L
sin
_
mπx
L
_
sin
_
nπx
L
_
dx =
_
L
−L
cos
_
mπx
L
_
cos
_
nπx
L
_
dx = 0,
and if n ≥ 1, then
_
L
−L
sin
2
_
nπx
L
_
dx =
_
L
−L
cos
2
_
nπx
L
_
dx = L.
With these integrals in mind, you may mimic the same procedure as before to
show that the coefﬁcients in a full Fourier series are given by
A
n
=
1
L
_
L
−L
φ(x) cos
_
nπx
L
_
dx (n = 0, 1, 2, . . .) (12.10)
B
n
=
1
L
_
L
−L
φ(x) sin
_
nπx
L
_
dx (n = 1, 2, 3, . . .). (12.11)
Observe that the scalar in front of these integrals is
1
L
, not
2
L
as with Fourier sine
and cosine coefﬁcients. Moreover, the interval of integration is [−L, L], not [0, L].
Example 12.1.10. Find the Fourier series representation of the discontinuous
function
φ(x) =
_
_
_
−2 if −3 < x < 0
6 if 0 ≤ x < 3.
on the interval −3 < x < 3. Solution: The fact that φ(x) has a jump discontinuity
at x = 0 is not a problem—when computing the Fourier coefﬁcients, we will
simply split the region of integration at the point where the discontinuity occurs.
Using (12.10)–(12.11) with L = 3, we compute
A
0
=
1
3
_
3
−3
φ(x) dx =
1
3
_
_
0
−3
−2 dx +
_
3
0
6 dx
_
= 4.
For n ≥ 1,
A
n
=
1
3
_
_
0
−3
−2 cos
_
nπx
3
_
dx +
_
3
0
6 cos
_
nπx
3
_
dx
_
= −
2
3
·
3
nπ
sin
_
nπx
3
_
¸
¸
¸
¸
0
−3
+ 2 ·
3
nπ
sin
_
nπx
3
_
¸
¸
¸
¸
3
0
.
344 convergence of fouri er seri es
Substituting in the limits of integration, we are left only with sines of integer
multiples of π. Therefore, A
n
= 0 for all n ≥ 1. The other coefﬁcients are given
by
B
n
=
1
3
_
_
0
−3
−2 sin
_
nπx
3
_
dx +
_
3
0
6 sin
_
nπx
3
_
dx
_
=
2
3
·
3
nπ
cos
_
nπx
3
_
¸
¸
¸
¸
0
−3
− 2 ·
3
nπ
cos
_
nπx
3
_
¸
¸
¸
¸
3
0
=
2
nπ
[1 −cos(−nπ)] −
6
nπ
[cos(nπ) −1]
=
2
nπ
[1 −(−1)
n
] −
6
nπ
[(−1)
n
−1] =
8
nπ
[1 −(−1)
n
] .
Equivalently,
B
n
=
_
_
_
16/nπ if n is odd
0 if n is even.
In summary, the only nonzero Fourier coefﬁcients are A
0
= 4 and B
n
for n odd.
The Fourier series for φ(x) on the interval −3 < x < 3 is given by
φ(x) = 2 +
16
π
sin
_
πx
3
_
+
16
3π
sin
_
3πx
3
_
+
16
5π
sin
_
5πx
3
_
+ · · ·
= 2 +
16
π
∞
∑
n=1
1
2n −1
sin
_
(2n −1)πx
3
_
.
Notice that, apart from the leading constant term A
0
/2, the only terms that are
present in this Fourier series are sine functions. Why are none of the cosine terms
present in this particular series? Why do some functions have series expansions
that require only sine terms, others have expansions with only cosine terms, and
some require a combination of both? We will address these questions in the next
section.
12.2. Convergence of Fourier Series
When we used the separation of variables technique to solve initialboundary
value problems for linear, homogeneous pdes, we presented the solutions as
inﬁnite series. In particular, this required the initial data to have Fourier series
representations. Fortunately, this is not a severe restriction at all—any physically
i ntroducti on to fouri er seri es 345
realistic choice of initial condition is guaranteed to have a convergent Fourier
series representation. In this section, we will explore classes of functions that have
convergent Fourier series representations. Virtually any “reasonable” function
φ(x) can be represented as the sum of its Fourier sine or cosine series on the
interval (0, L), or as the sum of its Fourier series on the interval (−L, L).
This last remark seems surprising, given that every term in a sine series is an
odd, periodic function, and every term in a cosine series is an even, periodic
function. Let us take a moment to review some properties of functions with
these types of symmetry. Recall that a function φ(x) is odd if φ(−x) = −φ(x) for
all x and φ(x) is even if φ(−x) = φ(x) for all x. A nonconstant function φ(x)
is called periodic if there exists a constant p > 0 such that φ(x + p) = φ(x) for
all x. The least positive p for which this statement holds is called the period of
φ(x). For example, the period of φ(x) = sin(3x) is p = 2π/3. Here are some
properties of odd, even and periodic functions:
If φ(x) is odd, then φ(0) = 0. To see why, note that φ(−0) = −φ(0).
If φ(x) is even, then φ(x) is automatically continuous at x = 0. If φ(x) also
differentiable at x = 0, then φ
/
(0) = 0. This follows from the limit deﬁnition of
the derivative, using centered differences:
φ
/
(0) = lim
h→0
φ(h) −φ(−h)
2h
= 0
since φ(h) = φ(−h).
If f (x) and g(x) are odd, then f (x) + g(x) is odd and f (x)g(x) is even. If
f (x) and g(x) are even, then both f (x) + g(x) and f (x)g(x) are even.
Suppose that f (x) is differentiable and integrable. If f (x) is odd then f
/
(x)
and
_
x
0
f (s)ds are even. If f (x) is even, then f
/
(x) and
_
x
0
f (s)ds are odd.
Let L > 0. Any function f (x) deﬁned on an interval [−L, L] can be written as
a sum of an odd function and an even function. Namely, f (x) = E(x) + O(x)
where
E(x) =
f (x) + f (−x)
2
and O(x) =
f (x) − f (−x)
2
.
It is easy to check that E(x) and O(x) are even and odd, respectively.
If φ(x) is odd on [−L, L] then
_
L
−L
φ(x)dx = 0.
The above list should convince you that odd and even functions obey considerably
different properties than odd and even numbers. For example, although the sum
346 convergence of fouri er seri es
of an odd number and an even number is always odd, the sum of an odd function
and an even function need not be odd or even.
When we discussed reﬂection methods for solving pdes with a single boundary
(i.e., on the domain [0, ∞)), we introduced the notion of the odd/even extensions
of functions. These concepts are deﬁned analogously on ﬁnite domains—suppose
that φ(x) is deﬁned on the interval [0, L]. The odd extension of φ(x) is deﬁned as
φ
odd
(x) =
_
¸
¸
¸
_
¸
¸
¸
_
φ(x) if 0 < x ≤ L
0 if x = 0
−φ(−x) if − L ≤ x < 0.
Likewise, the even extension of φ(x) is deﬁned as
φ
even
(x) =
_
_
_
φ(x) if 0 ≤ x ≤ L
φ(−x) if − L ≤ x < 0.
We introduce one more notion that will be useful in our discussion of (full)
Fourier series. Suppose φ(x) is deﬁned on some interval −L < x < L where
L > 0. Then the periodic extension of φ(x) is deﬁned as
φ
p
(x) = φ(x −2nL) if − L +2nL < x < L +2nL,
for all integers n. Notice that when n = 0, the deﬁnition of φ
p
reduces to that of
φ(x).
What do these remarks have to do with Fourier series? In the Fourier sine
series
φ(x) =
∞
∑
n=1
A
n
sin
_
nπx
L
_
(0 < x < L),
all of the terms in the series are odd and periodic. Since the frequencies of the
sine functions are positive integer multiples of frequency of the leading term, the
period of the sum must equal the period of the lowestfrequency term, which is
2L. What function does the series represent, assuming it actually converges? If
φ(x) is deﬁned on 0 < x < L, ﬁrst form its odd extension φ
odd
to the interval
−L < x < L, and then form the periodic extension of φ
odd
to the entire real line.
The sine series should represent that extended function on the entire real line.
i ntroducti on to fouri er seri es 347
Likewise, Fourier cosine series can be used to represent even, periodic functions
with period 2L.
When solving boundary value problems for pdes, we rarely concern ourselves
with the behavior of Fourier series representations of solutions outside the
physical domain, typically 0 ≤ x ≤ L. However, it is worth noting that the
types of homogeneous boundary conditions that we most commonly encounter
are related to the three types of extensions described above. Odd extensions
are associated with homogeneous Dirichlet conditions u(0, t) = 0 = u(L, t),
whereas even extensions are associated with homogeneous Neumann conditions
u
x
(0, t) = 0 = u
x
(L, t). Finally, if our domain is −L ≤ x < L, then the periodic
boundary conditions u(−L, t) = u(L, t) and u
x
(−L, t) = u
x
(L, t) are associated
with periodic extensions. Observe that the periodic extension of a function
satisfying these periodic boundary conditions is differentiable at every odd
integer multiple of L (i.e., at every junction between consecutive subintervals in
the periodic extension).
12.2.1 Norms, distances, inner products, and convergence. Our main goal for
the remainder of this chapter is to classify the types of functions that have
convergent Fourier series representations. In basic calculus, you learned what it
means for a series of numbers to converge to a number M—namely, the sequence
of partial sums must converge to M. It is less clear how to deﬁne what it means
for a series of functions (e.g., a Fourier series) to converge to a limiting function
over some interval. Indeed, there are many different notions of convergence
that we could adopt, some more natural than others. There are three types of
convergence that we shall introduce for sequences/series of functions. The main
difference between these types of convergence lies in how we measure distances
between functions.
Norms. Consider the class of continuous, realvalued functions deﬁned on the
interval [a, b]. There are many ways to quantify the “size” of a function f (x).
For example, if f (x) is continuous then  f (x) is guaranteed to achieve some
maximum value on the interval [a, b]. The quantity
f 
∞
= max
a≤x≤b
 f (x) (12.12)
348 convergence of fouri er seri es
is an example of a norm. Norms are used to measure the sizes of various
mathematical objects, such as vectors or functions. The norm deﬁned in (12.12)
above has several important features. First, it is clear that f 
∞
≥ 0 for all
continuous functions f (x) on [a, b] and that f 
∞
= 0 if and only if f (x) = 0
everywhere on [a, b]. Second, if α is any real constant, then αf 
∞
= αf 
∞
.
Third, if f and g are both continuous functions on [a, b], then f + g
∞
≤
f 
∞
+ g
∞
, the triangle inequality. In general, any norm (not just the one
deﬁned in (12.12)) must satisfy the three properties listed here.
The norm (12.12) deﬁned above is one of two important norms that we will
use from now on. The other norm is called the L
2
norm (read “Ltwo norm”).
For functions f deﬁned on [a, b], the L
2
norm is deﬁned as
f 
L
2 =
_
_
b
a
 f (x)
2
dx
_
1/2
, (12.13)
provided that the integral converges. Observe that every continuous function
on [a, b] certainly has a ﬁnite L
2
norm. However, the L
2
norm can be used to
measure the “size” of functions that are not necessarily continuous. To check
that (12.13) really is a norm, ﬁrst observe
1
that f 
L
2 ≥ 0 is clear from the
deﬁnition. Moreover, if α is any real constant, then it is easy to check that
αf 
L
2 = αf 
L
2 . Verifying the other properties requires considerably more
effort. In fact, proving the inequality f + g
L
2 ≤ f 
L
2 +g
L
2 , or equivalently
_
_
b
a
 f (x) + g(x)
2
dx
_
1/2
≤
_
_
b
a
 f (x)
2
dx
_
1/2
+
_
_
b
a
g(x)
2
dx
_
1/2
,
is surprisingly nontrivial. This special instance of the triangle inequality is
actually named in honor of the ﬁrst person to prove it: Minkowski’s inequality is
proved in most textbooks on mathematical analysis.
Distances. Once a set of functions is equipped with a norm, we automatically
inherit a natural notion of distance between functions. Namely, given two func
tions f and g, we can deﬁne the distance between f and g as the norm of the
difference f − g. Consider, for example, the two norms deﬁned above. If f (x)
and g(x) are continuous functions deﬁned on [a, b], one way of measuring their
1
Since f need not be continuous, it is actually not the case that f 
L
2 = 0 if and only if f is
zero everywhere. If f is zero except at ﬁnitely many points, then certainly f 
L
2 = 0 as well. This
seemingly subtle point is discussed in great detail in the mathematical subject of measure theory.
i ntroducti on to fouri er seri es 349
distance is to use
f −g
∞
= max
a≤x≤b
 f (x) −g(x).
Graphically, this crude notion of distance measures the maximum vertical gap
between the graphs of f and g over the interval [a, b]. It is sometimes called the
L
∞
distance between f and g. Alternatively, we could deﬁne the L
2
distance
between f and g as
f −g
L
2 =
_
_
b
a
[ f (x) −g(x)]
2
dx
_
1/2
.
The L
2
distance gives a better sense of the “cumulative” discrepancy between
the graphs of f and g. Anytime f (x) and g(x) deviate from one another, the
quantity [ f (x) −g(x)]
2
makes a positive contribution to the value of the integral.
The bigger the gap, the larger the contribution.
Example 12.2.1. Consider the functions f (x) = x and g(x) = x
2
on the interval
[0, 1]. Calculate the L
∞
and L
2
distances between f (x) and g(x) over this interval.
Solution: The L
∞
distance is given by
f −g
∞
= max
0≤x≤1
x −x
2
 = max
0≤x≤1
(x −x
2
),
where we have dropped the absolute value bars because x ≥ x
2
on the interval
[0, 1]. Finding the maximum value amounts to a calculus problem: the extreme
values of the continuous function x − x
2
must occur either (i) at one of the
endpoints of the interval [0, 1] or (ii) at an interior critical point. The derivative
of x −x
2
is 1 −2x, which is zero when x =
1
2
. Therefore, we must compare the
values of x −x
2
when x = 0, x =
1
2
, and x = 1. The maximum value of
1
4
occurs
at the interior critical point, and we conclude that
f −g
∞
=
1
4
.
As for the L
2
distance, we calculate
f −g
L
2 =
_
_
1
0
[x −x
2
]
2
dx
_
1/2
=
_
_
1
0
x
2
−2x
3
+ x
4
dx
_
1/2
350 convergence of fouri er seri es
=
_
x
3
3
−
x
4
2
+
x
5
5
¸
¸
¸
¸
1
0
_
1/2
=
_
1
30
.
As is typically the case, the L
∞
and L
2
distances are unequal, because they
measure distance between functions in completely different ways. If two contin
uous functions f and g have ﬁnite L
∞
distance on the interval [a, b], then they
automatically have ﬁnite L
2
distance (try to prove it). However, it is not hard to
create an example of two functions f and g with inﬁnite L
∞
distance but ﬁnite
L
2
distance!
The L
2
inner product. The set of functions on [a, b] with ﬁnite L
2
norm, denoted
by L
2
[a, b], has considerable structure. Algebraically, L
2
[a, b] is a vector space
whose elements are functions, and the various vector space axioms are easy
to check. Geometrically, L
2
[a, b] is actually endowed with an inner product: a
generalized version of the familiar dot product from multivariable calculus. This
enables us to deﬁne geometric notions such as orthogonality (perpendicularity)
of functions.
Deﬁnition 12.2.2. Suppose f (x) and g(x) are realvalued functions belonging to
the set L
2
[a, b]. Then the L
2
inner product of f and g is deﬁned as
¸ f , g¸ =
_
b
a
f (x)g(x) dx. (12.14)
If ¸ f , g¸ = 0, we say that f and g are orthogonal.
Notice that the inner product of two functions is a scalar. Unlike the norm of
a function, it is certainly possible for an inner product of two functions to be
negative.
Example 12.2.3. Consider the functions f (x) = x
2
, g(x) = −1 and h(x) = x
on the interval −2 ≤ x ≤ 2. Certainly all three of these functions belong to
L
2
[−2, 2] because continuous functions deﬁned on closed intervals are always
square integrable. The L
2
norm of f is given by
f 
L
2
[−2,2]
=
_
_
2
−2
x
2

2
dx
_
1/2
=
_
_
2
−2
x
4
dx
_
1/2
=
_
x
5
5
¸
¸
¸
¸
2
−2
dx
_
1/2
=
_
64
5
.
i ntroducti on to fouri er seri es 351
Similarly, you can check that the L
2
norm of g(x) is 2. The inner product of f
and g is
¸ f , g¸ =
_
2
−2
−x
2
dx = −
x
3
3
¸
¸
¸
¸
2
−2
= −
16
3
and the inner product of g and h is
¸g, h¸ =
_
2
−2
−x dx = 0.
We conclude that g and h are orthogonal in the space L
2
[−2, 2].
Many vector spaces other than L
2
are equipped with their own inner products.
For example, the familiar dot product is an inner product on threedimensional
Euclidean space R
3
. However, not every vector space can be endowed with an
inner product—only those with a certain level of geometric structure. There are
several properties that every inner product must satisfy. We list these properties
for the L
2
inner product: suppose that f and g are realvalued functions in the
space L
2
[a, b]. Then
¸ f , g¸ = ¸g, f ¸.
¸ f + g, h¸ = ¸ f , h¸ +¸g, h¸.
¸αf , g¸ = α¸ f , g¸ for all real constants α.
¸ f , f ¸ ≥ 0.
Notice that ¸ f , f ¸ is equal to the square of the L
2
norm of f .
Types of convergence. Recall that the norms (12.12) and (12.13) give us different
ways of measuring the distance between two functions f (x) and g(x) deﬁned on
an interval I. With notions of distance in mind, we are able to deﬁne what it
means for a sequence of functions {f
n
(x)}
∞
n=1
to converge to a limiting function
f (x) on I. The ﬁrst type of convergence that we shall introduce is called uniform
convergence:
Deﬁnition 12.2.4. Suppose that f
n
(x) is a sequence of functions deﬁned on an
interval I. We say that the sequence converges uniformly to f (x) on I if
f
n
(x) − f (x)
∞
= max
x∈I
 f
n
(x) − f (x) → 0 as n →∞.
Students who have completed courses in mathematical analysis will realize that
the maximum appearing in Deﬁnition 12.2.4 should technically be a supremum.
352 convergence of fouri er seri es
In the examples below, we generally work with sequences of continuous functions
over closed intervals I. That way, we are assured that the functions actually
achieve maximum values.
Example 12.2.5. Consider the sequence of functions f
n
(x) = x
n
, where n ≥ 1.
We claim that this sequence converges uniformly to f (x) = 0 on the interval
I = [0,
1
2
]. To see this, we calculate
f
n
(x) − f (x)
∞
= max
0≤x≤
1
2
x
n
−0 = max
0≤x≤
1
2
x
n
=
_
1
2
_
n
.
Since
_
1
2
_
n
→ 0 as n →∞, we have f
n
(x) − f (x)
∞
→ 0 as n →∞, as required.
Example 12.2.6. When testing the convergence of a sequence of functions, it is
important to specify the domain. In the previous example, suppose that we
had worked on the interval [0, 1) instead of [0,
1
2
]. We claim that the sequence
f
n
(x) = x
n
does not converge uniformly to f (x) = 0 on the interval [0, 1). To see
why, notice that no matter how large n is, there are always x values within the
interval [0, 1) for which f
n
(x) ≥
1
2
. Indeed, we have x
n
≥
1
2
whenever
n
_
1
2
≤ x < 1.
Regardless of n, it must be the case that f
n
(x) − f (x)
∞
≥
1
2
on the interval
[0, 1). Hence, f
n
(x) does not converge uniformly to f (x) = 0 on that interval.
There is a subtle point here worth mentioning. Above, we noted that no matter
how large n is, there will always be x values inside the interval [0, 1) for which
f
n
(x) ≥
1
2
. By contrast, suppose we ﬁx an x value and examine what happens
to f
n
(x) as n → ∞. Since 0 ≤ x < 1, it must be the case that x
n
→ 0 as n → ∞.
Therefore, for each ﬁxed x in the domain [0, 1), the sequence f
n
(x) converges to 0
as n →∞.
The remarks in the preceding example may seem surprising. We demonstrated
that it is possible to have a sequence of functions f
n
(x) that does not converge
uniformly to a function f (x) on an interval I even though f
n
(x) → f (x) for each
ﬁxed x ∈ I. In this case, the functions f
n
exhibit a somewhat “weaker” version of
convergence that we now deﬁne.
i ntroducti on to fouri er seri es 353
Deﬁnition 12.2.7. Suppose that f
n
(x) is a sequence of functions deﬁned on an
interval I. We say that the sequence converges pointwise to f (x) on I if
 f
n
(x) − f (x) → 0 as n →∞
for each x ∈ I.
Take a moment to compare the deﬁnitions of pointwise and uniform conver
gence. At ﬁrst glance, it may be difﬁcult to distinguish these two notions of
convergence. If a sequence of functions f
n
(x) converges uniformly to f (x) on
some interval I, then
max
x∈I
 f
n
(x) − f (x) → 0 as n →∞.
Since the maximum gap between the functions is approaching 0, then certainly
 f
n
(x) − f (x) → 0 as n →∞
for each ﬁxed x ∈ I. This argument proves that
Uniform convergence implies pointwise convergence.
On the other hand, the converse is not true, as illustrated by the example
f
n
(x) = x
n
on the interval I = [0, 1). That sequence of functions converges
pointwise to f (x) = 0 on I, but the convergence is not uniform.
Example 12.2.8. Let f
n
(x) =
1
n
sin(nπx/L) on the interval [0, L]. This sequence
of functions converges uniformly to f (x) = 0 on [0, L]. To see why, we exploit
the boundedness of the sine function to make the estimate
f
n
(x) − f (x)
∞
= max
0≤x≤L
¸
¸
¸
¸
1
n
sin
_
nπx
L
_
¸
¸
¸
¸
≤
1
n
.
Therefore, f
n
(x) − f (x)
∞
→ 0 as n →∞, which means that the sequence f
n
(x)
converges uniformly to the function f (x) = 0. Since the sequence converges
uniformly, it automatically converges pointwise as well.
Example 12.2.9. Consider the sequence of functions
f
n
(x) =
x
n
1 + x
n
n = 1, 2, 3, . . .
354 convergence of fouri er seri es
on the interval I = [0, 2]. To test for pointwise convergence, we will calculate
lim
n→∞
f
n
(x) for each ﬁxed x ∈ I. If 0 ≤ x < 1, then x
n
→ 0 as n → ∞, and it
follows that
lim
n→∞
f
n
(x) = lim
n→∞
x
n
1 + x
n
= 0 if 0 ≤ x < 1.
If x = 1, then f
n
(x) =
1
2
for all n, and therefore lim
n→∞
f
n
(1) =
1
2
. Finally, note
that if 1 < x ≤ 2, then x
n
→∞ as n →∞. Using L’Hôpital’s rule, we calculate
lim
n→∞
f
n
(x) = lim
n→∞
x
n
1 + x
n
= lim
n→∞
nx
n−1
nx
n−1
= 1 if 1 < x ≤ 2.
Putting everything together, we have shown that the sequence f
n
(x) converges
pointwise to the discontinuous function
f (x) =
_
¸
¸
¸
_
¸
¸
¸
_
0 if 0 ≤ x < 1
1
2
if x = 1
1 if 1 < x ≤ 2.
The convergence is not uniform—can you explain why?
Now that we are more accustomed to pointwise and uniform convergence, we
introduce a third type of convergence:
Deﬁnition 12.2.10. Suppose that f
n
(x) is a sequence of functions deﬁned on an
interval I. We say that the sequence converges to f (x) in the L
2
sense on I if
f
n
(x) − f (x)
L
2 → 0 as n →∞.
If I = [a, b], stating that f
n
(x) → f (x) in the L
2
sense is equivalent to saying
that
_
b
a
[ f
n
(x) − f (x)]
2
dx → 0 as n →∞.
Example 12.2.11. In a previous example, we showed that the sequence f
n
(x) =
x
n
converges pointwise to f (x) = 0 on the interval [0, 1), but that the convergence
is not uniform. Does the sequence converge to f (x) = 0 in the L
2
sense? We
i ntroducti on to fouri er seri es 355
calculate
f
n
(x) − f (x)
2
L
2
=
_
1
0
[x
n
−0]
2
dx =
_
1
0
x
2n
dx =
x
2n+1
2n +1
¸
¸
¸
¸
1
0
=
1
2n +1
.
Letting n →∞, we see that f
n
(x) − f (x)
2
L
2
→ 0 as n →∞. It follows that the
sequence f
n
(x) does converge to f (x) = 0 in the L
2
sense on [0, 1).
Above, we noted that uniform convergence is “stronger” than pointwise
convergence: if f
n
(x) → f (x) uniformly on an interval I, then we automatically
conclude that f
n
(x) → f (x) pointwise on I as well. A similar statement holds
with respect to L
2
convergence, namely,
Uniform convergence on a ﬁnite interval implies convergence in the L
2
sense.
To see why, suppose that f
n
(x) → f (x) uniformly on [a, b], and let
M
n
= f
n
(x) − f (x)
∞
= max
a≤x≤b
 f
n
(x) − f (x).
Our assumption of uniform convergence is equivalent to saying that M
n
→ 0 as
n →∞. To test for convergence in the L
2
sense, observe that
f
n
(x) − f (x)
L
2 =
_
_
b
a
[ f
n
(x) − f (x)]
2
dx
_
1/2
≤
_
_
b
a
M
2
n
dx
_
1/2
=
_
M
2
n
(b −a) = M
n
√
b −a.
Since M
n
→ 0 as n →∞, it follows that f
n
(x) − f (x)
L
2 → 0 as well. We have
now proved that f
n
(x) → f (x) in the L
2
sense.
Example 12.2.12. For sequences of functions deﬁned over inﬁnite domains, uni
form convergence need not imply convergence in the L
2
sense. Consider the
sequence of functions deﬁned by
f
n
(x) =
_
_
_
1
√
n
if 0 ≤ x ≤ n
0 if x > n.
This sequence converges uniformly to f (x) = 0 on the inﬁnite domain 0 ≤ x < ∞
because
f
n
(x) − f (x)
∞
= max
0≤x<∞
 f
n
(x) − f (x) =
1
√
n
356 convergence of fouri er seri es
approaches 0 as n → ∞. However, the sequence {f
n
(x)} does not converge to
f (x) in the L
2
sense because
f
n
(x) − f (x)
L
2 =
_
_
∞
0
[ f
n
(x) − f (x)]
2
dx
_
1/2
=
_
_
n
0
1
n
dx
_
1/2
= 1
for all n. Since f
n
(x) − f (x)
L
2 → 1 ,= 0 as n →∞, we do not have convergence
in the L
2
sense.
In our study of Fourier series, we are interested in pointwise, uniform, and L
2
convergence of series of functions, not sequences. Fortunately, the deﬁnitions of
these various modes of convergence can be easily extended. Given an inﬁnite
series of functions
∞
∑
n=1
f
n
(x), (12.15)
the Nth partial sum is deﬁned as
S
N
(x) =
N
∑
n=1
f
n
(x).
Deﬁnition 12.2.13. We say that the series (12.15) converges to a function f (x)
pointwise on (a, b) if  f (x) −S
N
(x) → 0 as N →∞ for each x ∈ (a, b),
uniformly on (a, b) if f (x) −S
N
(x)
∞
→ 0 as N →∞,
in the L
2
sense on (a, b) if f (x) −S
N
(x)
L
2 → 0 as N →∞.
Again, uniform convergence is the strongest of these three types of con
vergence: if a series converges uniformly on the ﬁnite interval (a, b), then it
automatically converges pointwise and in the L
2
sense.
Example 12.2.14. In calculus, you studied the geometric series
∞
∑
n=0
x
n
.
We claim that this series converges pointwise but not uniformly to f (x) =
1/(1 −x) on the interval (−1, 1). To verify pointwise convergence, consider the
partial sum
S
N
(x) =
N
∑
n=0
x
n
,
i ntroducti on to fouri er seri es 357
which is a polynomial of degree N. This summation simpliﬁes nicely if we use
an algebra trick: notice that
(1 −x)S
N
(x) = (1 −x)(1 + x + x
2
+ · · · + x
N
) = 1 −x
N+1
.
Therefore, if x ,= 1,
S
N
(x) =
1 −x
N+1
1 −x
.
Testing for pointwise convergence, we measure
 f (x) −S
N
(x) =
¸
¸
¸
¸
1
1 −x
−
1 −x
N+1
1 −x
¸
¸
¸
¸
=
¸
¸
¸
¸
x
N+1
1 −x
¸
¸
¸
¸
.
Assuming that −1 < x < 1, notice that x
N+1
→ 0 as N →∞. Therefore,
 f (x) −S
N
(x) → 0 as n →∞
for each ﬁxed x ∈ (−1, 1). This is precisely what we needed in order to conclude
that our series converges pointwise.
To show that the series does not converge uniformly, we must show
2
that
f (x) −S
N
(x)
∞
= max
−1<x<1
 f (x) −S
N
(x)
does not converge to 0 as N →∞. It sufﬁces to prove that for each choice of N,
there are always x values in the interval (−1, 1) for which  f (x) −S
N
(x) ≥ 1.
From our proof of pointwise convergence, we already know that
f (x) −S
N
(x) =
x
N+1
1 −x
,
so let us try to show that the equation
x
N+1
1 −x
= 1
always has a solution x ∈ (−1, 1) regardless of N. The equation can be rewritten
as x
N+1
+ x −1 = 0. Letting p(x) = x
N+1
+ x −1 denote the lefthand side,
2
A word of caution here. On open intervals such as −1 < x < 1, even continuous functions
are not guaranteed to achieve a maximum/minimum value. Students who have taken a course in
mathematical analysis will realize that we should really be taking a supremum instead of a maximum.
358 convergence of fouri er seri es
notice that p(x) is continuous. Moreover, observe that p(0) = −1 < 0 and
p(1) = 1 > 0 regardless of N. Since p(x) is continuous and transitions from
negative to positive as x varies from 0 to 1, the Intermediate Value Theorem from
calculus tells us that p(x) must have a root on the interval (0, 1). This is precisely
what we needed to show, and we conclude that the series does not converge
uniformly on the interval (−1, 1).
Determining whether the series converges in the L
2
sense is more challenging.
According to the deﬁnition of L
2
convergence, we would need to evaluate
lim
N→∞
f (x) −S
N
(x)
L
2
and check whether the limit is zero. This requires that we integrate the square of
f (x) −S
n
(x)—i.e.,
_
1
−1
[ f (x) −S
N
(x)]
2
dx =
_
1
−1
x
2N+2
(1 −x)
2
dx.
Notice that the integral is improper and should therefore be written as a limit
lim
b→1
−
_
b
−1
x
2N+2
(1 −x)
2
dx.
This integral seems a bit tricky—one way to evaluate it is to substitute x = u +1,
use the binomial theorem to expand the numerator, and integrate one term at a
time. It turns out that the integral diverges regardless of N, from which it follows
that the series we started with does not converge in the L
2
sense. Thankfully, we
will soon state a much cleaner test for L
2
convergence of series.
Example 12.2.15. For nonnegative integers n, deﬁne the functions f
n
(x) =
x
n
−x
n+2
on the interval 0 < x < 1. We claim that the inﬁnite series
∞
∑
n=0
f
n
(x)
converges to f (x) = 1 + x both pointwise and in the L
2
sense on (0, 1), but not
uniformly. The partial sums collapse nicely because this is a telescoping series:
S
N
(x) =
N
∑
n=0
f
n
(x) = (1 −x
2
) + (x −x
3
) + (x
2
−x
4
) + · · ·
i ntroducti on to fouri er seri es 359
+ (x
N−1
−x
N+1
) + (x
N
−x
N+2
) = 1 + x −x
N+1
−x
N+2
.
Testing for pointwise convergence, notice that
 f (x) −S
N
(x) =
¸
¸
¸x
N+1
+ x
N+2
¸
¸
¸ = x
N+1
(1 + x) → 0 as N →∞,
for 0 < x < 1. This establishes that f
n
(x) → f (x) pointwise on (0, 1). To prove
L
2
convergence, we calculate
f (x) −S
N
(x)
L
2 =
_
_
1
0
[ f (x) −S
N
(x)]
2
dx
_
1/2
=
_
_
1
0
[x
N+1
+ x
N+2
]
2
dx
_
1/2
=
_
_
1
0
x
2N+2
+2x
2N+3
+ x
2N+4
dx
_
1/2
=
_
x
2N+3
2N +3
+
2x
2N+4
2N +4
+
x
2N+5
2N +5
¸
¸
¸
¸
1
0
_
1/2
=
_
1
2N +3
+
2
2N +4
+
1
2N +5
_
1/2
→ 0 as N →∞.
Finally, to see that the convergence is not uniform, it sufﬁces to note that the
maximum value of
 f (x) −S
N
(x) = x
N+1
+ x
N+2

on the closed interval [0, 1] is 2, and is achieved at the right endpoint of that inter
val. Since x
N+1
−x
N+2
is a continuous function, there must be x values in the
open interval (0, 1) for which x
N+1
−x
N+2
is arbitrarily close to 2. Consequently,
it is impossible for f (x) −S
N
(x)
∞
to converge to 0 as N →∞, and therefore
the series does not converge uniformly to f (x) on (0, 1).
12.2.2 Convergence theorems. Armed with our understanding of the three
different notions of convergence that we have singled out, we now state theorems
regarding convergence of Fourier series.
360 convergence of fouri er seri es
Theorem 12.2.16. Convergence in the L
2
sense. Suppose that φ(x) is deﬁned on
an interval [a, b]. If φ(x)
L
2
[a,b]
is ﬁnite, then Fourier series for φ(x) converges
to φ(x) in the L
2
sense on (a, b).
Example 12.2.17. In a previous example, we showed that the Fourier series for
φ(x) =
_
_
_
−2 if −3 < x < 0,
6 if 0 ≤ x < 3.
on the interval −3 < x < 3 is given by
2 +
16
π
∞
∑
n=1
1
2n −1
sin
_
(2n −1)πx
3
_
.
To verify that the series really does converge to the function φ(x) in the L
2
sense,
we must check that φ(x)
L
2 < ∞. Indeed,
φ(x)
L
2 =
_
_
3
−3
φ(x)
2
dx
_
1/2
=
_
_
0
−3
(−2)
2
dx +
_
3
0
6
2
dx
_
1/2
< ∞.
The exact value of the integral is irrelevant—the fact that it is ﬁnite assures that
the Fourier series for φ(x) converges to φ(x).
Not surprisingly, since uniform convergence is stronger than convergence in
the L
2
sense, the class of functions φ(x) whose Fourier series converge uniformly
is considerably smaller:
Theorem 12.2.18. Uniform convergence. Suppose that φ(x) is deﬁned and
continuous on an interval [−L, L] and φ(−L) = φ(L). If φ
/
(x) is piecewise
continuous (with only jump discontinuities) on [−L, L], then the Fourier series
for φ(x) converges to φ(x) uniformly.
There are variants of Theorem 12.2.18 that guarantee uniform convergence
under slightly different conditions. Notice that this theorem would not guarantee
uniform convergence of the Fourier series in the preceding example, because the
function was discontinuous.
Example 12.2.19. Theorem 12.2.18 guarantees that the Fourier series for φ(x) =
x on the interval −10 ≤ x ≤ 10 converges uniformly. To verify that the
conditions of the theorem are satisﬁed, ﬁrst observe that φ(x) is continuous on
[−10, 10] and φ(−10) = φ(10). Also, φ
/
(x) = −1 for x < 0 and φ
/
(x) = 1 for
i ntroducti on to fouri er seri es 361
x > 0. Therefore, the only discontinuity of φ
/
(x) occurs at x = 0 and is a jump
discontinuity.
Finally, we turn our attention to pointwise convergence of Fourier series. The
criteria we shall state for pointwise convergence involve the notions of left and
righthand limits and derivatives. The left and righthand limits of f (x) at x
0
are
deﬁned as
f (x
−
0
) = lim
x→x
−
0
f (x) and f (x
+
0
) = lim
x→x
+
0
f (x),
respectively, provided that these onesided limits exist. If f (x
−
0
) and f (x
+
0
) exist
but are unequal, we say that f (x) has a jump discontinuity at x
0
. Furthermore, if
f (x
−
0
) = f (x
0
) = f (x
+
0
), then f (x) is continuous at x
0
. The left and righthand
derivatives of f (x) at x
0
are deﬁned as
f
/
(x
−
0
) = lim
h→0
+
f (x
−
0
) − f (x
0
−h)
h
and f
/
(x
+
0
) = lim
h→0
+
f (x
0
+ h) − f (x
+
0
)
h
respectively, provided that these onesided limits exist.
Theorem 12.2.20 (Pointwise convergence.). Suppose that φ(x) is deﬁned and
piecewise continuous (with only jump discontinuities) on an interval [−L, L]. If
the left and righthand derivatives of φ(x) exist at each jump discontinuity, then
the Fourier series for φ(x) converges pointwise to
1
2
[φ(x
+
) + φ(x
−
)] for each
x ∈ (−L, L).
Here are some remarks that may help you interpret this theorem:
Notice that
1
2
[φ(x
+
) +φ(x
−
)] represents the average value of the left and right
hand limits of φ(x). If φ(x) is continuous at x = x
0
, then φ(x
−
0
) = φ(x) = φ(x
+
0
),
in which case the expression
1
2
[φ(x
+
) + φ(x
−
)] reduces to φ(x). In other words,
at points where φ(x) is continuous, its Fourier series is guaranteed to converge
pointwise to φ(x) itself.
To determine what happens at the endpoints x = ±L, examine the peri
odic extension of φ(x). If the righthand derivative φ
/
(−L
+
) and the lefthand
derivative φ
/
(L
−
) exist, then the Fourier series for φ(x) converges pointwise to
1
2
[φ(−L) + φ(L)] at both x = ±L.
Following up on the preceding two remarks, if φ(x) is continuous and
φ(−L) = φ(L), then the Fourier series for φ(x) automatically converges point
wise to φ(x) on the entire interval [−L, L].
362 convergence of fouri er seri es
Example 12.2.21. Consider the piecewise continuous function
φ(x) =
_
¸
¸
¸
_
¸
¸
¸
_
−2 if −8 < x < −3
3 if −3 ≤ x ≤ 2
7 if 2 < x < 8.
This function satisﬁes the conditions of the pointwise convergence theorem (The
orem 12.2.20). Namely, φ(x) is continuous, and the only types of discontinuities
are jump discontinuities. The Fourier series for φ(x) must converge pointwise to
φ(x) itself everywhere except at the jumps. More exactly, the Fourier series for
φ(x) would converge pointwise to
˜
φ(x) =
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
−2 if −8 < x < −3
1
2
if x = −3
3 if −3 ≤ x ≤ 2
5 if x = 2
7 if 2 < x < 8.
Notice that at x = −3, the series converges to the average of the left and right
hand limits:
1
2
[φ(−3
+
) + φ(−3
−
)] =
1
2
. Similarly, at x = 2, the Fourier series
converges to
1
2
[3 +7] = 5.
At the endpoints x = ±8, the series would converge pointwise to the average
of f (−8
+
) = −2 and f (8
−
) = 7. That is, the Fourier series would converge to
5
2
when x = ±8.
We remark that the theorems regarding pointwise, uniform and L
2
convergence
of Fourier series on [−L, L] are easily adapted to handle Fourier sine and cosine
series on an interval [0, L]. If (i) φ(x) is continuous on [0, L] with φ(0) = φ(L)
and (ii) φ
/
(x) is piecewise continuous on [0, L] (any discontinuities are jump
discontinuities), then the Fourier sine & cosine series for φ(x) converge uniformly
to φ(x) on [0, L]. On a similar note, if φ(x) has a ﬁnite L
2
norm over the
interval [0, L], then the Fourier sine & cosine series for φ(x) converge to φ(x)
in the L
2
sense. Although the pointwise convergence theorem 12.2.20 has a
natural counterpart for sine and cosine series, our earlier remarks concerning
convergence at the endpoints x = 0 and x = L are no longer valid. After all, a
i ntroducti on to fouri er seri es 363
Fourier sine series on [0, L] must converge to zero at both endpoints, independent
of how the function φ(x) is deﬁned at those two points.
Exercises
1. Suppose m and n are nonnegative integers and let L > 0. Show that
_
L
0
cos
_
mπx
L
_
cos
_
nπx
L
_
dx =
_
¸
¸
¸
_
¸
¸
¸
_
L if m = n = 0
L/2 if m = n ,= 0
0 if m ,= n.
You may wish to use the following trigonometric identities:
cos α cos β =
1
2
cos(α − β) +
1
2
cos(α + β) and cos
2
α =
1
2
[1 +cos(2α)] .
2. Solve the homogeneous Dirichlet problem
u
tt
= c
2
u
xx
(0 ≤ x ≤ 5)
u(x, 0) = sin
_
17πx
5
_
u
t
(x, 0) = 0
u(0, t) = 0
u(5, t) = 0.
Feel free to quote (11.28), the general solution of the wave equation on a ﬁnite
interval with homogeneous Dirichlet boundary conditions. Describe how the
solution behaves, as well as a physical interpretation.
3. Calculate the Fourier sine series for φ(x) = x on the interval 0 < x < 3 and
plot the ﬁrst three partial sums of the series.
4. Calculate the Fourier cosine series for the discontinuous function
φ(x) =
_
_
_
1 if 0 < x <
5
2
3 if
5
2
≤ x < 5.
on the interval 0 < x < 5.
364 convergence of fouri er seri es
5. Calculate the (full) Fourier series for φ(x) = x on the interval −4 < x < 4.
6. Show that for any choice of nonnegative integers m and n, the functions
f (x) = sin(mπx/L) and g(x) = cos(nπx/L) on the interval [−L, L] are
orthogonal with respect to the L
2
inner product.
7. Consider the set of continuous functions on the interval [a, b]. The L
1
norm of
such functions is deﬁned as
f 
L
1
=
_
b
a
 f (x) dx.
Show that this really is a norm, by verifying that (i) f 
L
1
≥ 0, with equality
if and only if f = 0; (ii) αf 
L
1
= αf 
L
1
for any real constant α; and (iii)
f + g
L
1
≤ f 
L
1
+g
L
1
for all functions f and g that are continuous on
[a, b].
8. By L
1
[a, b], we mean the set of integrable functions on [a, b] which have ﬁnite
L
1
norm (see previous problem). We say that a sequence of functions {f
n
}
∞
n=1
converges to f in the L
1
sense if
_
b
a
 f (x) − f
n
(x) dx → 0 as n →∞.
Consider the functions
f
n
(x) =
_
_
_
1
n
if 0 ≤ x ≤ n
0 if n < x
deﬁned on the interval [a, b] = [0, ∞). Show that f
n
(x) → f (x) = 0 pointwise
as n →∞. Then show that the sequence f
n
(x) does not converge to f (x) = 0
in the L
1
sense.
9. Consider the functions f
n
(x) = x
−n
on the interval [1, 2]. Does this sequence
of functions converge pointwise as n →∞? If so, what function f (x) does the
sequence converge to? Does the sequence converge uniformly?
10. The purpose of this exercise is to show that pointwise convergence need not
imply convergence in the L
2
sense. Consider the functions
f
n
(x) =
_
_
_
n
2
(1 −nx) if 0 < x ≤
1
n
0 if
1
n
< x < 1.
i ntroducti on to fouri er seri es 365
Show that this sequence converges pointwise to some function f (x) (which
you will need to ﬁnd) on the interval 0 < x < 1. Then, show that the sequence
does not converge to f (x) in the L
2
sense.
11. The purpose of this exercise is to show that convergence in the L
2
sense
need not imply pointwise convergence. Consider the piecewise continuous
functions
f
n
(x) =
_
_
_
1 if
1
2
−
1
n
≤ x ≤
1
2
+
1
n
0 otherwise
on the interval 0 ≤ x ≤ 1. Show that this sequence converges in the L
2
sense
to f (x) = 0, but that the sequence does NOT converge pointwise to f (x).
12. Solve the Dirichlet problem
u
t
+2u = u
xx
(0 < x < 1)
u(x, 0) = φ(x) =
_
_
_
x if 0 ≤ x ≤
1
2
1 −x if
1
2
< x ≤ 1,
u(0, t) = u(1, t) = 0
expressing your solution in terms of a Fourier sine series.
13. Let φ(x) be deﬁned as in the previous exercise. Explain why Fourier sine se
ries expansion of φ(x) converges uniformly to φ(x) on [0, 1]. Does it converge
in the L
2
sense?
14. These questions concern the function
φ(x) =
_
¸
¸
¸
_
¸
¸
¸
_
0 if −9 < x ≤ −3,
2 if −3 < x ≤ 3,
0 if 3 < x < 9.
(a) Without computing the Fourier series for φ(x), determine whether the
series converges pointwise on the interval (−9, 9). If so, what does the
series converge to?
(b) Without computing the Fourier series for φ(x), determine whether the
series converges in the L
2
sense on the interval (−9, 9).
(c) Compute the (full) Fourier series representation of φ(x) on the interval
(−9, 9).
366 convergence of fouri er seri es
(d) Let S
3
(x) denote the sum of the ﬁrst three nonzero terms in the (full)
Fourier series representation you calculated in Part (c). Carefully plot
S
3
(x) and φ(x) on the same set of axes. Either use a computer to generate
the plot for you, or give an accurate handsketch.
CHAPTER 13
The Laplace and Poisson Equations
U
p to now, we have dealt almost exclusively with pdes for which one
independent variable corresponds to time. Now, we will analyze a pde
for which this is not the case: Laplace’s equation. To motivate where Laplace’s
equation comes from and what it models, consider the heat equation with
periodic boundary conditions:
u
t
= κu
xx
(−L ≤ x ≤ L)
u(−L, t) = u(L, t)
u
x
(−L, t) = u
x
(L, t).
Physically, this system models diffusion of heat within a thin “onedimensional”
ring of wire. You can imagine “bending” the interval −L ≤ x ≤ L into a ring.
With the two ends x = −L and x = L in contact, they correspond to the same
physical location (hence the two boundary conditions).
Using the separation of variables technique, you can show that the solution of
the above system is given by
u(x, t) =
A
0
2
+
∞
∑
n=1
_
A
n
cos
_
nπx
L
_
+ B
n
sin
_
nπx
L
__
e
−n
2
π
2
κt/L
2
.
Given an initial condition u(x, 0) = φ(x) for −L ≤ x ≤ L, the constants appear
ing in the formula for u(x, t) are given by the Fourier coefﬁcients
A
n
=
1
L
_
L
−L
φ(x) cos
_
nπx
L
_
dx and B
n
=
1
L
_
L
−L
φ(x) sin
_
nπx
L
_
dx.
367
368
Question: How do solutions behave as t →∞? Intuitively, we would expect the
exponential factor in the above series to decay rapidly to 0 as t →∞. This would
suggest that only the leading term survives as t →∞:
lim
t→∞
u(x, t) =
A
0
2
=
1
2L
_
L
−L
φ(x) dx
for all x ∈ [−L, L]. Notice that this integral represents the average value of the
initial temperature distribution φ(x). Consequently, we expect the solution of
the heat equation on a ringshaped domain (periodic boundary conditions) to
approach a steadystate in which heat is uniformly distributed around the ring.
Moving beyond the above example, let us ask a more general question: What
are the steadystate solutions of the heat equation (or the wave equation)? In
other words, are there solutions the heat equation that are timeindependent?
Physical intuition suggests that if we apply a ﬁxed temperature distribution along
the boundary of the spatial domain, then the heat proﬁle should equilibrate to
some steady conﬁguration throughout the domain as t →∞.
Example 13.0.22. In one spatial dimension, ﬁnding steadystate solutions of
the heat equation is rather easy. Suppose that we wish to ﬁnd steadystate
solutions of u
t
= κu
xx
on the domain 0 < x < L, subject to homogeneous
Dirichlet boundary conditions u(0, t) = 0 = u(L, t). Since such solutions are
timeindependent, we should set u
t
= 0, reducing the pde to an ode κu
//
(x) = 0.
Integrating twice, we ﬁnd that u(x) = C
1
x + C
2
, where C
1
and C
2
are constants.
The boundary conditions imply that both of the constants are zero, which means
that the only steadystate solution of this Dirichlet problem is the constant
function u = 0. As an exercise, ﬁnd the steadystate solution of the heat equation
with the more general Dirichlet conditions u(0, t) = τ
1
and u(L, t) = τ
2
. You
should ﬁnd that temperature varies linearly between τ
1
and τ
2
over the domain
0 ≤ x ≤ L.
A similar argument shows that any constant function is a steadystate solution
of the onedimensional heat equation with homogeneous Neumann boundary
conditions.
In higher dimensions, it is more challenging to ﬁnd the steadystate solutions
of the heat equation on a given spatial domain Ω. In two or three dimensions,
the heat equation takes the form u
t
= κ(u
xx
+ u
yy
) or u
t
= κ(u
xx
+ u
yy
+ u
zz
),
respectively, where κ > 0 is the diffusion coefﬁcient. These equations can be
the laplace and poi sson equati ons 369
written more compactly if we introduce the Laplace operator (or Laplacian)
∆ =
_
∂
2
∂x
2
+
∂
2
∂y
2
_
in two space dimensions
∆ =
_
∂
2
∂x
2
+
∂
2
∂y
2
+
∂
2
∂z
2
_
in three space dimensions.
(13.1)
The heat equation now takes the form u
t
= κ∆u, and it should always be
clear from the context as to whether we are working in two or three spatial
dimensions. Likewise, the two and threedimensional wave equations can be
written as u
tt
= c
2
∆u.
Steadystate solutions of the heat equation are timeindependent, which im
plies that u
t
= 0. Setting u
t
= 0 in the heat equation leads to an important pde
that gets a special name:
Deﬁnition 13.0.23. The pde
∆u(x, y) =
∂
2
u
∂x
2
+
∂
2
u
∂y
2
= 0
is called the twodimensional Laplace equation. The pde
∆u(x, y, z) =
∂
2
u
∂x
2
+
∂
2
u
∂y
2
+
∂
2
u
∂z
2
= 0
is called the threedimensional Laplace equation. Solutions of the Laplace equation
are called harmonic functions.
For emphasis, solutions of Laplace’s equation can be thought of as solutions
of the “steadystate heat equation”, because Laplace’s equation can be obtained
by setting u
t
= 0 in u
t
= κ∆u. One way to interpret this physically is to imagine
a pie that is placed in an oven of constant temperature. The oven temperature
imposes a Dirichlet condition on the boundary (external surface) of the pie. As
time passes, the temperature within the pie equilibrates to some steadystate
distribution. This steadystate temperature proﬁle will be a harmonic function—a
solution to Laplace’s equation on a pieshaped domain.
Laplace’s equation arises in a variety of physical contexts, not just steady
state heat distribution. For example, the velocity potential of an incompressible,
irrotational ﬂuid satisﬁes Laplace’s equation. It is also worth noting that solutions
370 di rchlet and neumann problems
of Laplace’s equation can be thought of as solutions of the “steadystate wave
equation” u
tt
= c
2
∆u = 0.
The inhomogeneous version of Laplace’s equation is also famous enough to
receive a special name:
Deﬁnition 13.0.24. The pde
∆u(x, y) =
∂
2
u
∂x
2
+
∂
2
u
∂y
2
= f (x, y)
is called the twodimensional Poisson equation. The pde
∆u(x, y, z) =
∂
2
u
∂x
2
+
∂
2
u
∂y
2
+
∂
2
u
∂z
2
= f (x, y, z)
is called the threedimensional Poisson equation.
Perhaps the most wellknown application of Poisson’s equation arises in
electrostatics. If ϕ(x, y, z) denotes the electric potential at a point (x, y, z) in
a static (timeindependent) electric ﬁeld, then Gauss’s Law states that ∆ϕ =
f (x, y, z), where f is proportional to the total charge density. In other words,
electrostatic potential satisﬁes Poisson’s equation.
13.1. Dirchlet and Neumann Problems
Our main goal for this chapter is to solve Laplace’s equation on certain bounded
domains Ω. In twodimensions, we will focus on domains Ω whose boundaries
are formed by a simple, closed, piecewise smooth curve. By simple, we mean
that the curve is not allowed to intersect itself, and by piecewise smooth, we
mean that the curve can be subdivided into ﬁnitely many curves, each of which
is inﬁnitely differentiable. We denote the boundary of such domains by ∂Ω—in
this context, the use of the ∂ symbol is a notational standard, and has nothing
to do with differentiation. Sketches of the types of domains we have in mind
appear in Figure 13.1.
To further describe the domains Ω we have in mind, let us crudely deﬁne a
few important topological notions. A set Ω is called open if it contains none
of its boundary points. In twodimensions, the set S of points (x, y) such that
x
2
+ y
2
< 1 forms an open set (a circular disc). The boundary ∂S would be the
the laplace and poi sson equati ons 371
Ω
Ω
Ω
Figure 13.1. Three examples of twodimensional domains Ω whose
boundaries ∂Ω consist of simple, closed, piecewise smooth curves.
circle x
2
+ y
2
= 1, none of which are contained in S. A set Ω in R
2
is called
bounded if there exists a positive number R (possibly very large) such that Ω is
contained within a disc or radius R. Finally, by a connected set Ω, we have in
mind a domain that consists of one contiguous “piece”. For example, on a map
of the United States, the state of Kansas is connected whereas the state of Hawaii
is disconnected since it is composed of multiple islands. Readers interested in
precise, technical deﬁnitions of openness, boundedness, and connectedness are
encouraged to complete a course in topology. The Dirichlet problem for Laplace’s
equation is formulated as followed. Let Ω be a two or threedimensional domain
of the type described above. Then the Dirichlet problem is given by
∆u(x, y) = 0 inside Ω
u(x, y) = h(x, y) on the boundary, ∂Ω
in two dimensions, or
∆u(x, y, z) = 0 inside Ω
u(x, y, z) = h(x, y, z) on the boundary, ∂Ω
in three dimensions. As explained previously, the Dirichlet boundary condition
prescribes the temperature distribution on ∂Ω. For example, if h(x, y) = x
2
+ y
2
,
then points on ∂Ω that are far from the origin will be hotter than points that are
close to the origin.
The Neumann problem for Laplace’s equation is a bit more difﬁcult to state,
because Neumann conditions describe heat ﬂux across the boundary, not the
actual temperature distribution. In two dimensions, suppose that (x
0
, y
0
) is a
372 well posedness and the maxi mum pri nci ple
point on the boundary ∂Ω and let n denote the outward unit normal vector at
that point. Recall that the normal derivative of u at (x
0
, y
0
) is deﬁned as
∂u
∂n
= ∇u(x
0
, y
0
) • n.
(This deﬁnition is easily extended to three or more dimensions.) The Neumann
problem for Laplace’s equation is formulated by specifying the normal derivative
of u at all points of the boundary; i.e.,
∆u(x, y) = 0 inside Ω
∂u
∂n
= h(x, y) on the boundary, ∂Ω
in two dimensions, or
∆u(x, y, z) = 0 inside Ω
∂u
∂n
= h(x, y, z) on the boundary, ∂Ω
in three dimensions. In this case, the function h describes the net outward
heat ﬂux at all points on the boundary. If the boundary is insulated to prevent
entry/exit of heat across the boundary ∂Ω, then h = 0.
13.2. Wellposedness and the Maximum Principle
Assuming that there actually exists a solution of the Dirichlet problem for
Laplace’s equation, we will prove that the solution is unique and is not sensitive to
small changes in the boundary condition. The proof makes use of the Maximum
Principle for Laplace’s equation, which we now state:
Theorem 13.2.1 (Maximum Principle for Laplace’s Equation). Let Ω be a set that
is bounded, open, and connected in either two or three dimensions, and let u
be a function is harmonic inside Ω and continuous on Ω∪ ∂Ω. If u attains its
maximum or minimum value inside Ω, then u is a constant function.
In other words, if u is a nonconstant solution of Laplace’s equation, then the
only place u can attain its maximum and minimum values is on ∂Ω, the boundary
of the domain. The proof is similar to that of the Maximum Principle for the
heat equation. The idea is to consider what would happen if u had a maximum
the laplace and poi sson equati ons 373
(or minimum) inside Ω. Then the second derivative test would require that both
u
xx
≤ 0 and u
yy
≤ 0. If either of these were strictly negative, then we would have
u
xx
+ u
yy
< 0, which contradicts our assumption that u is a solution of Laplace’s
equation u
xx
+ u
yy
= 0. Therefore, we may assume without loss of generality
that u
xx
= u
yy
= 0 at any maximum/minimum occurring inside Ω. Explaining
why this implies that u would have to be constant throughout Ω requires care,
and the interested reader is encouraged to consult a more comprehensive text on
pdes (such as Strauss [10]).
Example 13.2.2. The function u(x, y) = x
2
−y
2
+ 3 is harmonic: u
xx
+ u
yy
= 0.
Suppose that we wish to ﬁnd the maximum and minimum values of this function
on the discshaped domain x
2
+ y
2
≤ 1. By the Maximum Principle, the extreme
values of u(x, y) must occur on the boundary of the domain, which is the circle
x
2
+ y
2
= 1. One way to maximize u(x, y) = x
2
−y
2
+3 is to use the method of
Lagrange multipliers from multivariable calculus. However, in this example we
can take a simpler approach: the constraint x
2
+ y
2
= 1 tells us that y
2
= 1 −x
2
.
Substituting this into the expression for u(x, y), we ﬁnd that u(x, y) = 2x
2
+ 2
for points (x, y) on the boundary of our disc. Maximizing this function of a
single variable is straightforward: the minimum of 2x
2
+2 occurs when x = 0.
The corresponding yvalues on the circular boundary are y = ±1. It follows that
the minimum value of u(x, y) is u(0, 1) = u(0, −1) = 2. Similarly, the maximum
value of u(x, y) is u(1, 0) = u(−1, 0) = 4.
Now let us use the Maximum Principle to prove that the Dirichlet problem for
Poisson’s equation
∆u = f inside Ω
u = h on ∂Ω
has a unique solution. (Note that Laplace’s equation is a special case of Poisson’s
equation with f = 0.) Suppose that u and v are both solutions to this Dirichlet
problem and deﬁne w = u −v. Then by linearity of the Laplace operator,
∆w = ∆(u −v) = ∆u −∆v = f − f = 0.
374 well posedness and the maxi mum pri nci ple
On the boundary, we ﬁnd that w = h − h = 0. It follows that w satisﬁes a
homogeneous Dirichlet problem for Laplace’s equation:
∆w = 0 inside Ω
w = 0 on ∂Ω.
By the Maximum Principle, both the maximum and minimum values of w are
attained on the boundary of the domain. This implies that w = 0 throughout
Ω∪ ∂Ω, from which we conclude that u = v. Hence, the solution of the Dirichlet
problem for Poisson’s equation (and therefore Laplace’s equation) is unique.
To test for stability with respect to small changes in the boundary condition,
let us compare the solutions of the Dirichlet problems
∆u = f inside Ω
u = g on ∂Ω
and
∆v = f inside Ω
v = h on ∂Ω.
In other words, u and v satisfy the same pde but different boundary conditions.
As before, we measure the gap between the solutions by deﬁning w = u −v.
Then w satisﬁes a Dirichlet problem for Laplace’s equation:
∆w = 0 inside Ω
w = g −h on ∂Ω.
By the Maximum Principle,
max
Ω∪∂Ω
u −v = max
Ω∪∂Ω
w = max
∂Ω
w = max
∂Ω
g −h.
In words, this says that the maximum separation between u and v throughout
the domain Ω∪ ∂Ω is at most as large as the maximum separation between g
and h on the boundary. This is precisely what we need in order to prove stability:
small discrepancies in boundary conditions (i.e., g and h are “close”) cannot
cause major changes in how the solutions u and v behave.
the laplace and poi sson equati ons 375
Now we can rest assured that the Dirichlet problem for Poisson’s equation
is wellposed, provided that we can actually ﬁnd a solution. We now turn our
attention to this task.
13.3. Translation and Rotation Invariance
Understanding a few “geometric” properties of Laplace’s equation will give us
intuition regarding what sorts of solutions we should seek. Take a moment to
compare what follows with the approach we took when solving the Cauchy
problem for the heat equation, (10.16)–(10.17).
Like the heat equation, Laplace’s equation is translation invariant as we shall
illustrate in two dimensions. If we make the change of coordinates ξ = x +α and
η = y + β, this has the effect of translating (shifting) every point in the plane in
the direction of the vector (α, β). Laplace’s equation u
xx
+ u
yy
= 0 can be written
in terms of the new coordinates (ξ, η) by using the chain rule:
u
ξ
= u
x
dx
dξ
+ u
y
dy
dξ
= u
x
u
η
= u
x
dx
dη
+ u
y
dy
dη
= u
y
,
and calculating the second derivatives yields u
xx
= u
ξξ
and u
yy
= u
ηη
. There
fore u
ξξ
+ u
ηη
= 0, demonstrating that Laplace’s equation is not affected by
translating the coordinates.
More interestingly, Laplace’s equation is also rotation invariant: making a
change of coordinates that rotates the plane through an angle θ (see Figure 13.2)
does not affect Laplace’s equation. In two dimensions, such a change of coordi
nates takes the form
_
ξ
η
_
=
_
cos θ −sin θ
sin θ cos θ
_ _
x
y
_
=
_
x cos θ −y sin θ
x sin θ + y cos θ
_
.
Verifying that this coordinate transformation really does rotate the plane through
an angle θ is not difﬁcult. You should check that (i) the vectors (x, y) and (ξ, η)
have the same length and (ii) the angle between these vectors is θ (the dot product
will help you see this). As before, we use the chain rule to write u
x
and u
y
in
376 translati on and rotati on i nvari ance
x
y
ξ
η
θ
Figure 13.2. Laplace’s equation is invariant under rotations of the
coordinate system.
terms of the new coordinates:
u
x
= u
ξ
dξ
dx
+ u
η
dη
dx
= (cos θ)u
ξ
+ (sin θ)u
η
u
y
= u
ξ
dξ
dy
+ u
η
dη
dy
= (−sin θ)u
ξ
+ (cos θ)u
η
.
Calculating the second partial derivatives, we ﬁnd that
u
xx
+ u
yy
= (sin
2
θ +cos
2
θ)(u
ξξ
+ u
ηη
) = u
ξξ
+ u
ηη
,
and it follows that rotating our coordinate system has no effect on Laplace’s
equation. We remark that Laplace’s equation is also translation and rotation
invariant in higher dimensions. Rotation invariance suggests that Laplace’s
equation may have solutions with radial symmetry. In two dimensions, let us
seek solutions of ∆u(x, y) = 0 of the form u(x, y) = v(r), where r =
_
x
2
+ y
2
measures distance from the origin. Since v is a function of one variable, we
expect that substituting u(x, y) = v(r) into Laplace’s equation will reduce the
pde to an ode. First, notice that
∂r
∂x
=
2x
2
_
x
2
+ y
2
=
x
r
and
∂r
∂y
=
2y
2
_
x
2
+ y
2
=
y
r
.
Now, use the chain rule to convert u
xx
+ u
yy
= 0 into an ode for v(r): the ﬁrst
partial derivatives of u become
u
x
=
∂
∂x
v(r) = v
/
(r)
∂r
∂x
=
x
r
v
/
(r)
the laplace and poi sson equati ons 377
u
y
=
∂
∂y
v(r) = v
/
(r)
∂r
∂y
=
y
r
v
/
(r).
Since r depends upon both x and y, calculating the second partial derivatives
requires careful use of the product/quotient and chain rules:
u
xx
=
∂
∂x
_
x ·
1
r
· v
/
(r)
_
=
1
r
v
/
(r) +
_
−
x
r
2
_ _
x
r
_
v
/
(r) +
_
x
r
_
2
v
//
(r),
or equivalently
u
xx
=
_
1
r
−
x
2
r
3
_
v
/
(r) +
_
x
r
_
2
v
//
(r).
Similarly,
u
yy
=
_
1
r
−
y
2
r
3
_
v
/
(r) +
_
y
r
_
2
v
//
(r).
Adding these two expressions,
0 = u
xx
+ u
yy
=
_
2
r
−
x
2
+ y
2
r
3
_
v
/
(r) +
_
x
2
+ y
2
r
2
_
v
//
(r).
Finally, since x
2
+ y
2
= r
2
, we have shown that radially symmetric solutions of
the twodimensional Laplace equation would have to obey the ode
v
//
(r) +
1
r
v
/
(r) = 0. (13.2)
Equation (13.2) is a secondorder, linear ode, but does not have constant coefﬁ
cients (which prevents us from solving it via the characteristic equation approach).
There are several ways to solve this ode—for example, we could reduce the order
by deﬁning w = v
/
, use separation of variables to solve the resulting ﬁrstorder
ode for w, and then integrate w to obtain v. Another approach is as follows: use
algebra to rewrite (13.2) as
v
//
(r)
v
/
(r)
= −
1
r
,
and observe that
v
//
(r)
v
/
(r)
=
d
dr
ln[v
/
(r)].
The ode becomes
d
dr
ln[v
/
(r)] = −
1
r
,
378 translati on and rotati on i nvari ance
u(x,y)
x
y
Figure 13.3. Graph of a radially symmetric solution (13.3) Laplace’s
equation u
xx
+ u
yy
= 0.
and integrating both sides yields
ln[v
/
(r)] = −ln(r) + C = ln
_
1
r
_
+ C,
where C is a constant of integration. Exponentiating both sides,
v
/
(r) = e
ln(1/r)+C
=
A
r
,
where A = e
C
is a constant. Finally, integrating a second time reveals that the
radially symmetric solutions of the twodimensional Laplace equation are given
by
v(r) = Alnr + B, (13.3)
where A and B are constants and r > 0. Equivalently, we have shown that
u(x, y) = Aln
_
x
2
+ y
2
+ B
are harmonic functions in the plane. A sketch of one such function is given
in Figure 13.3. The special case A = −
1
2π
and B = 0 is sometimes called
the fundamental solution of the Laplace equation in two dimensions. In three
dimensions, the fundamental solution of Laplace’s equation ∆u(x, y, z) = 0 has
a considerably different form. As before, let us seek solutions u(x, y, z) = v(r),
where r =
_
x
2
+ y
2
+ z
2
. Recycling our previous calculations,
u
x
=
x
r
v
/
(r) u
y
=
y
r
v
/
(r) and u
z
=
z
r
v
/
(r),
the laplace and poi sson equati ons 379
and
u
xx
=
_
1
r
−
x
2
r
3
_
v
/
(r) +
_
x
r
_
2
v
//
(r)
u
yy
=
_
1
r
−
y
2
r
3
_
v
/
(r) +
_
y
r
_
2
v
//
(r)
u
zz
=
_
1
r
−
z
2
r
3
_
v
/
(r) +
_
z
r
_
2
v
//
(r).
Adding these yields
0 = u
xx
+ u
yy
+ u
zz
=
_
3
r
−
x
2
+ y
2
+ z
2
r
3
_
v
/
(r) +
_
x
2
+ y
2
+ z
2
r
2
_
v
//
(r),
and since x
2
+ y
2
+ z
2
= r
2
, we obtain the ode
v
//
(r) +
2
r
v
/
(r) = 0. (13.4)
Despite the apparent similarity between Equations (13.2) and (13.4), the solu
tions are quite different. Fortunately, the solution technique is identical: ﬁrst,
write (13.4) as
d
dr
ln[v
/
(r)] = −
2
r
.
Integrating both sides with respect to r,
ln[v
/
(r)] = −2 lnr + C = ln
_
1
r
2
_
+ C,
where C is a constant of integration. Exponentiating both sides,
v
/
(r) = e
ln(1/r
2
)+C
=
A
r
2
,
where A = e
C
is a constant. Finally, integrating a second time reveals that the
radially symmetric solutions of the threedimensional Laplace equation are given
by
v(r) = −
A
r
+ B, (13.5)
380 translati on and rotati on i nvari ance
where A and B are constants and r > 0. Equivalently, we have shown that
u(x, y, z) = −
A
_
x
2
+ y
2
+ z
2
+ B (x, y, z) ,= (0, 0, 0)
are harmonic functions in threedimensional space. The special case A = −1
and B = 0 is sometimes called the fundamental solution of the Laplace equation in
three dimensions.
Example 13.3.1. Solve Laplace’s equation on the annular (ringshaped) domain
1 < r < e, with Dirichlet boundary conditions u = 8 when r = 1 and u = 12
when r = e. Solution: Due to the symmetry of the domain, it is natural to work
from (13.3) instead of using Cartesian coordinates. The two boundary conditions
tell us that
8 = Aln(1) + B = B and 12 = Aln(e) + B = A + B.
It follows that A = 4 and B = 8, and the solution is given by v(r) = 4 lnr +8.
Equivalently,
u(x, y) = 4 ln
_
x
2
+ y
2
+8 = 2 ln(x
2
+ y
2
) +8.
By wellposedness of the Dirichlet problem, we know that this is the unique
solution of Laplace’s equation on the given domain.
In order to interpret the solution physically, you may wish to graph u(x, y) over
the annular domain. The boundary of the domain consists of two circles, and the
Dirichlet boundary conditions tell us that the temperature at the “inner boundary”
is held constant at 8 while the temperature at the outer boundary is held constant
at 12. If we travel radially outward within the domain (from the inner boundary
toward the outer boundary) the temperature increases logarithmically. This may
seem counterintuitive, since one might guess that temperature would increase
linearly from 8 to 12 as we move radially outward.
Example 13.3.2. When we sought radial solutions u(x, y) = v(r) of the Laplace
equation in two dimensions, we calculated that
u
xx
+ u
yy
= v
//
(r) +
1
r
v
/
(r) (r ,= 0).
the laplace and poi sson equati ons 381
Multiplying both sides by r, notice that
r(u
xx
+ u
yy
) = rv
//
(r) + v
/
(r) =
d
dr
_
rv
/
(r)
¸
.
This observation can be useful when solving the Laplace (or Poisson) equations
on annular domains. For example, let us solve the Poisson equation u
xx
+u
yy
= 1
on the domain 3 < r < 5, with Dirichlet boundary conditions u = 10 on the
circle r = 3 and u = 0 on the circle r = 5. Multiplying both sides of Poisson’s
equation by r, we have
r(u
xx
+ u
yy
) = r
and, from the above remarks, we conclude that
d
dr
_
rv
/
(r)
¸
= r.
Integrating both sides with respect to r,
rv
/
(r) =
r
2
2
+ A,
where A is a constant of integration. Dividing both sides by r and integrating a
second time,
u(x, y) = v(r) =
r
2
4
+ Alnr + B.
Remark. Since Poisson’s equation is nothing more than an inhomogeneous
version of Laplace’s equation, it is not surprising that this formula for v(r)
contains the terms Alnr + B, which we recognize from formula (13.3).
Finally, we will use the two Dirichlet conditions to solve for the constants A
and B:
10 = v(3) =
9
4
+ Aln3 + B
0 = v(5) =
25
4
+ Aln5 + B.
This is a system of two equations in two unknowns:
Aln3 + B =
31
4
382 translati on and rotati on i nvari ance
Aln5 + B = −
25
4
,
which has solution
A =
14
ln(3/5)
and B =
31
4
−
14 ln3
ln(3/5)
.
Inserting these into the above formula for u(x, y), the solution of our Dirichlet
problem is given by
u(x, y) =
x
2
+ y
2
4
+
14 ln
_
x
2
+ y
2
ln(3/5)
+
31
4
−
14 ln3
ln(3/5)
.
Example 13.3.3. Solve the mixed DirichletNeumann problem
∆u(x, y, z) =
_
x
2
+ y
2
+ z
2
inside 1 < r < 2
u = 12 if r = 1
∂u
∂n
= 3 if r = 2.
Solution: The domain is a spherical shell. Recall that when we sought solutions
u(x, y, z) = v(r) of the Laplace equation in three dimensions, we calculated that
∆u = u
xx
+ u
yy
+ u
zz
= v
//
(r) +
2
r
v
/
(r) (r ,= 0).
If we multiply both sides of this equation by r
2
, observe that
r
2
∆u = r
2
v
//
(r) +2rv
/
(r) =
d
dr
[r
2
v
/
(r)].
In the present example, multiplying our Poisson equation ∆u = r by r
2
yields
d
dr
[r
2
v
/
(r)] = r
3
.
Integrating with respect to r yields
r
2
v
/
(r) =
r
4
4
+ A,
the laplace and poi sson equati ons 383
where A is a constant of integration. Divide by r
2
and integrate a second time to
get
v(r) =
r
3
12
−
A
r
+ B.
We recognize the latter two terms from the formula (13.5) for radial solutions
of Laplace’s equation in three dimensions. (Again, this is not surprising since
Poisson’s equation is the inhomogeneous Laplace equation.)
The Dirichlet condition on the inner boundary (r = 1) tells us that v(1) = 12.
The Neumann boundary condition speciﬁes the net outward ﬂux of heat at the
outer boundary r = 2, and the vector n points radially outward in the direction of
increasing r. In terms of v, the Neumann boundary condition becomes v
/
(2) = 3.
Solving for A and B, you will ﬁnd that A = 8 and B = 239/12. In terms of the
original coordinates, the overall solution is
u(x, y, z) =
_
x
2
+ y
2
+ z
2
_
3/2
12
−
8
_
x
2
+ y
2
+ z
2
+
239
12
.
Remark. If a Neumann condition had been given at the inner boundary, we must
exercise caution when writing the boundary condition in terms of our variable
v(r). Suppose, for example, that our boundary condition on the inner shell r = 1
had been the Neumann condition ∂u/∂n = 12. On the shell r = 1, the outward
normal vector n is directed towards the origin in the direction of decreasing r. Hence,
in terms of v, the boundary condition would be v
/
(1) = −12, not v
/
(1) = 12.
13.4. Laplace’s Equation on Bounded Domains
In the previous section, we found all radial solutions of Laplace’s equation on
unbounded domains: two and threedimensional Euclidean space. Then, we used
our results to solve a few boundary value problems on domains with radial
symmetry (annuli and spherical shells). Finding explicit solutions of Laplace’s
equation on general bounded domains is usually too much to expect. In this
section, we will solve Dirichlet problems for Laplace’s equation on two very
special twodimensional domains: rectangles and discs.
13.4.1 Dirichlet problem on a rectangle. Let Ω denote the open rectangular
domain 0 < x < a and 0 < y < b in the plane. Our goal is to solve the Dirichlet
384 laplace’ s equati on on bounded domai ns
y
x
b
0
0
a
u = B(x)
u = T(x)
u
=
L
(
y
)
Ω
u
=
R
(
y
)
Figure 13.4. Illustration of the rectangular domain Ω with Dirichlet
boundary conditions speciﬁed on each edge.
problem
∆u(x, y) = 0 inside Ω
u(x, 0) = B(x) 0 < x < a
u(x, b) = T(x) 0 < x < a
u(0, y) = L(y) 0 < y < b
u(a, y) = R(y) 0 < y < b.
(13.6)
The functions B(x), T(x), L(y), and R(y) specify temperature distributions
on the (B)ottom, (T)op, (L)eft, and (R)ight edges of the rectangle, respectively.
Figure 13.4 illustrates this domain as well as the given boundary conditions. We
can solve the system (13.6) by splitting this problem into four simpler Dirichlet
problems in which three of the four edges are given homogeneous Dirichlet
conditions. More precisely, suppose that u(x, y) denotes the solution of the above
Dirichlet problem with boundary conditions (B, T, L, R). Let u
B
(x, y) denote the
solution of the Dirichlet problem with boundary conditions (B, 0, 0, 0)—i.e., let
T(x) = L(y) = R(y) = 0 in (13.6) but leave the condition on the bottom edge
of the rectangle unchanged. Similarly, let u
T
, u
L
and u
R
denote the solution
of the Dirichlet problem with boundary conditions (0, T, 0, 0), (0, 0, L, 0), and
(0, 0, 0, R), respectively. Then the solution of the general Dirichlet problem (13.6)
is obtained by summing the solutions of the four special Dirichlet problems:
u = u
B
+ u
T
+ u
L
+ u
R
.
the laplace and poi sson equati ons 385
Let us solve one of these four simpler Dirichlet problems, in which only the
top edge of the rectangle receives a nonhomogeneous condition:
∆u(x, y) = 0 inside Ω
u(x, 0) = 0 0 < x < a
u(x, b) = T(x) 0 < x < a
u(0, y) = 0 0 < y < b
u(a, y) = 0 0 < y < b.
(13.7)
We will solve (13.7) via separation of variables—suppose that u(x, y) = X(x)Y(y).
Laplace’s equation becomes X
//
Y + XY
//
= 0 and, by algebra,
X
//
X
= −
Y
//
Y
.
Since the lefthand side is a function of x only whereas the righthand side is a
function of y only, the only way this can happen is if both expressions are equal
to a common constant, say −λ. The result is a system of two odes: X
//
+ λX = 0
and Y
//
− λY = 0. We will solve the X equation ﬁrst, because its boundary
conditions are easy to incorporate. Since u(x, y) = X(x)Y(y), the boundary
conditions at the left and right edges of the rectangle tell us that
0 = u(0, y) = X(0)Y(y) and 0 = u(a, y) = X(a)Y(y).
We exclude the possibility that Y(y) = 0, since that would imply that u(x, y) = 0,
which would not satisfy the boundary condition on the top edge of the rectangle.
Therefore, the boundary conditions for our X equation are X(0) = 0 = X(a). We
have encountered the twopoint boundary value problem
X
//
+ λX = 0 X(0) = 0 = X(a)
on numerous occasions. For λ ≤ 0, this problem has no nonzero solutions,
implying that all eigenvalues are positive. The eigenvalues are λ
n
= (nπ/a)
2
for
n ≥ 1, and corresponding eigenfunctions are
X
n
(x) = sin
_
nπx
a
_
(n ≥ 1).
386 laplace’ s equati on on bounded domai ns
With these eigenvalues in mind, we turn our attention to the differential equation
for Y(y), namely Y
//
−λY = 0. Let Y
n
(y) denote the solution of this equation
corresponding to the eigenvalue λ = λ
n
. The characteristic equation is m
2
−
(nπ/a)
2
= 0, which has distinct, real roots ±nπ/a. Therefore,
Y
n
(y) = E
n
e
nπy/a
+ F
n
e
−nπy/a
,
where E
n
and F
n
are constants. The boundary condition on the lower edge of the
rectangular domain tells us that Y
n
(0) = 0, but it is less clear how to incorporate
the boundary condition on the top edge of the rectangle. Ignoring the upper
boundary for the moment, the condition Y
n
(0) = 0 indicates that E
n
+ F
n
= 0,
from which
Y
n
(y) = E
n
_
e
nπy/a
− e
−nπy/a
_
(n ≥ 1).
Each of the functions
u
n
(x, y) = X
n
(x)Y
n
(y) = E
n
_
e
nπy/a
− e
−nπy/a
_
sin
_
nπx
a
_
satisﬁes Laplace’s equation as well as the Dirichlet boundary conditions on every
edge of the rectangle except for the top one. Taking an inﬁnite linear combination
of these solutions, we ﬁnd that
u(x, y) =
∞
∑
n=1
G
n
u
n
(x, y) =
∞
∑
n=1
E
n
G
n
_
e
nπy/a
− e
−nπy/a
_
sin
_
nπx
a
_
is the general solution of the Laplace equation subject to the homogeneous
Dirichlet conditions on the left, right, and bottom edges of the rectangular
domain. The constants can be combined by introducing A
n
= E
n
G
n
so that our
general solution takes the form
u(x, y) =
∞
∑
n=1
A
n
_
e
nπy/a
− e
−nπy/a
_
sin
_
nπx
a
_
.
It remains to incorporate the boundary condition on the upper edge of the
rectangle, u(x, b) = T(x). Setting y = b in our series solution, we ﬁnd that
T(x) =
∞
∑
n=1
A
n
_
e
nπb/a
− e
−nπb/a
_
sin
_
nπx
a
_
.
the laplace and poi sson equati ons 387
Because we have ﬁxed y = b, notice that the two exponential terms are now
independent of the two variables x and y. Hence, we may absorb these terms
into our coefﬁcients A
n
by introducing
˜
A
n
= A
n
_
e
nπb/a
− e
−nπb/a
_
.
In doing so, the function T(x) takes the form of a Fourier sine series
T(x) =
∞
∑
n=1
˜
A
n
sin
_
nπx
a
_
.
The Fourier coefﬁcients are given by
˜
A
n
=
2
a
_
a
0
T(x) sin
_
nπx
a
_
dx
and, consequently,
A
n
=
2
a
_
e
nπb/a
− e
−nπb/a
_
−1
_
a
0
T(x) sin
_
nπx
a
_
dx.
Inserting these Fourier coefﬁcients into our general formula for u(x, y), we have
shown that the solution of (13.7) is given by
u(x, y) =
2
a
∞
∑
n=1
e
nπy/a
− e
−nπy/a
e
nπb/a
− e
−nπb/a
_
_
a
0
T(x) sin
_
nπx
a
_
dx
_
sin
_
nπx
a
_
. (13.8)
Formula (13.8) may seem complicated, but it is not too difﬁcult to interpret if
we make a few observations about the various terms/factors in the series. Notice
that as y increases from 0 to b, the factor
e
nπy/a
− e
−nπy/a
e
nπb/a
− e
−nπb/a
increases from 0 to 1. On the other hand, the other factors
2
a
_
_
a
0
T(x) sin
_
nπx
a
_
dx
_
sin
_
nπx
a
_
are nothing more than terms in a Fourier sine series representation for T(x).
Combining these two observations, the effect of the exponential factors should be
to “damp” the function T(x) as y decreases from b towards 0. This is illustrated
388 laplace’ s equati on on bounded domai ns
x
y
u
(
x
,
y
)
T(x)
Figure 13.5. Solution of (13.7) for a particular choice of T(x).
in Figure (13.5) for a particular choice of temperature distribution T(x) applied
at the top edge of the rectangle. Recall that (13.8) is merely the solution of one out
of four Dirichlet problems that we must solve to build the solution of the general
Dirichlet problem (13.6). Fortunately, we may recycle many of our calculations to
solve the other Dirichlet problems. For example, suppose that we wish to solve
the Dirichlet problem
∆u(x, y) = 0 inside Ω
u(x, 0) = 0 0 < x < a
u(x, b) = 0 0 < x < a
u(0, y) = 0 0 < y < b
u(a, y) = R(y) 0 < y < b,
(13.9)
in which only the right edge of the rectangle receives a nonhomogeneous
boundary condition. This is essentially identical to the problem (13.7), except
that the roles of x and y are reversed and R(y) now plays the role of T(x).
Retracing our steps, it is convenient to separate the variables as
−
X
//
X
=
Y
//
Y
= −λ,
and solve the Y equation ﬁrst because its boundary conditions will be homoge
neous: Y(0) = 0 = Y(b). Afterward, solving (13.9) is literally a matter of quoting
the laplace and poi sson equati ons 389
formula (13.8) with (i) the variables x and y swapped; (ii) the constants a and b
swapped; and (iii) the Fourier sine series for T(x) replaced with the Fourier sine
series for R(y).
Remark. The same solution procedure works for other types of boundary condi
tions as well. We could impose Neumann boundary conditions along each edge
of the rectangular domain, or we could mix and match Dirichlet and Neumann
conditions along the different edges of the rectangle.
We can also use the separation of variables technique to solve Laplace’s
equation on a boxshaped domain in three dimensions. For example, let Ω
denote the domain 0 < x < a, 0 < y < b, and 0 < z < c, and consider the
Dirichlet problem
∆u(x, y, z) = 0 inside Ω
u(x, y, c) = T(x, y) top face of ∂Ω
u = 0 on the rest of ∂Ω.
As usual, we seek separated solutions of the form u(x, y, z) = X(x)Y(y)Z(z) so
that Laplace’s equation becomes X
//
YZ + XY
//
Z + XYZ
//
= 0. Dividing by XYZ,
we have
X
//
X
+
Y
//
Y
+
Z
//
Z
= 0.
The three ratios appearing in this sum must equal a common constant −λ (ex
plain). In particular, the X equation satisﬁes the eigenvalue problem X
//
+λX = 0
with X(0) = 0 = X(a). The eigenvalues and corresponding eigenfunctions are
given by
λ
n
=
_
nπ
a
_
2
and X
n
(x) = sin
_
nπx
a
_
, (n ≥ 1).
The Y equation satisﬁes essentially the same eigenvalue problem, leading us to
deﬁne
Y
m
(y) = sin
_
nπy
b
_
, (m ≥ 1).
Rewriting the Z equation as
Z
//
Z
= −
X
//
X
−
Y
//
Y
,
390 laplace’ s equati on on bounded domai ns
we seek solutions in which X = X
n
and Y = Y
m
. Since
X
//
n
X
n
= −
(nπ/a)
2
sin(nπx/a)
sin(nπx/a)
= −
_
nπ
a
_
2
and
Y
//
m
Y
m
= −
(mπ/a)
2
sin(mπx/a)
sin(mπx/a)
= −
_
mπ
a
_
2
,
we must solve
Z
//
m,n
=
_
_
nπ
a
_
2
+
_
mπ
a
_
2
_
Z
m,n
(m, n ≥ 1).
One of the boundary conditions on Z is Z(0) = 0, which is easy to implement.
However, the key observation regarding this ode for Z is that it is doubly indexed.
When we build the general solution of this Dirichlet problem, we must sum over
both m and n, leading to a double sum. Ultimately, we would need to develop a
theory of Fourier series in several variables and, although this is a straightforward
extension of our efforts in the previous chapter, we will not deal with such series
here.
13.4.2 Dirichlet problem on a disc. The fundamental solution of Laplace’s
equation in two dimensions has radial symmetry throughout the plane, so we
may suspect that ∆u(x, y) = 0 has “nice” solutions on (bounded) discshaped
domains centered at the origin. In this section, we will solve a classic problem in
pdes: Find all bounded solutions of the Dirichlet problem
u
xx
+ u
yy
= 0 if x
2
+ y
2
< a
2
u = h(θ) if x
2
+ y
2
= a
2
.
(13.10)
Here, a > 0 denotes the radius of the discshaped domain, and θ denotes an angle
used to parametrize the boundary (0 ≤ θ < 2π). The function h(θ) speciﬁes a
temperature distribution applied along the boundary of the disc. Problem (13.10)
is famous enough in the study of pdes that it is simply referred to as the Dirichlet
problem on a disc. (When mathematicians mention this problem, it is understood
that the underlying pde is Laplace’s equation.)
We will solve (13.10) by ﬁrst writing Laplace’s equation in polar coordinates
(r, θ), and then applying the separation of variables technique. Let x = r cos θ
the laplace and poi sson equati ons 391
and y = r sin θ and note that x
2
+ y
2
= r
2
and θ = arctan(y/x). By implicit
differentiation, we calculate
∂r
∂x
=
x
r
,
∂r
∂y
=
y
r
,
∂θ
∂x
= −
y
r
2
, and
∂θ
∂y
=
x
r
2
.
Now by the chain rule,
u
x
=
∂u
∂r
∂r
∂x
+
∂u
∂θ
∂θ
∂x
=
x
r
u
r
−
y
r
2
u
θ
and
u
y
=
∂u
∂r
∂r
∂y
+
∂u
∂θ
∂θ
∂y
=
y
r
u
r
+
x
r
2
u
θ
.
Tedious use of the chain rule allows us to express the second partial derivatives
u
xx
and u
yy
in terms of u
rr
and u
θθ
, and it turns out that
u
xx
+ u
yy
= u
rr
+
1
r
u
r
+
1
r
2
u
θθ
.
Now that we know how to write the Laplacian in polar coordinates, the Dirichlet
problem for a disc can be expressed as
u
rr
+
1
r
u
r
+
1
r
2
u
θθ
= 0 if r < a
u = h(θ) if r = a.
(13.11)
Our goal is to seek bounded solutions of (13.11). The boundedness will become
important later when we attempt to solve a secondorder ode in the variable r
despite having only one boundary condition.
Equation (13.11) can be solved by separation of variables: let u(r, θ) =
R(r)Θ(θ). Laplace’s equation becomes
R
//
Θ+
1
r
R
/
Θ+
1
r
2
RΘ
//
= 0
which, after dividing by RΘ and multiplying by r
2
, yields
r
2
R
//
+ rR
/
R
+
Θ
//
Θ
= 0.
392 laplace’ s equati on on bounded domai ns
With the variables thus separated, we conclude that
Θ
//
Θ
= −
r
2
R
//
+ rR
/
R
= −λ,
a constant. The result is a system of two secondorder odes
Θ
//
+ λΘ = 0
r
2
R
//
+ rR
/
−λR = 0.
At ﬁrst, it seems unclear how to set up and solve an eigenvalue problem for Θ.
However, if we recall that Θ represents an angular variable, we may enforce a
periodic boundary condition
Θ(θ +2π) = Θ(θ) for all choices of θ.
It is easy to check that if λ < 0, then Θ
//
+ λΘ = 0 has no nontrivial solutions
satisfying this periodic boundary condition. On the other hand, λ = 0 is an
eigenvalue. In that case, the ode reduces to Θ
//
= 0, implying that Θ(θ) = Cθ +D
where C and D are constants. The boundary condition Θ(θ +2π) = Θ(θ) implies
that C = 0, but there are no restrictions on D. As for the positive eigenvalues, let
λ = β
2
where β > 0. The general solution of the ode for Θ is
Θ(θ) = Ccos(βθ) + Dsin(βθ),
where C and D are constants. The only way that the boundary condition
Θ(θ +2π) = Θ(θ) can be satisﬁed for all choices of θ is if both
cos(βθ +2πβ) = cos(βθ) and sin(βθ +2πβ) = sin(βθ) for all θ.
Thus, 2πβ must be a positive
1
integer multiple of 2π, implying that β can be
any positive integer. Deﬁning β
n
= n, the eigenvalues for the Θ equation are
λ
n
= β
2
n
= n
2
for n ≥ 0. The eigenfunctions corresponding to these eigenvalues
have the form
Θ
0
(θ) = C
0
1
Recall that β was assumed positive.
the laplace and poi sson equati ons 393
and
Θ
n
(θ) = C
n
cos(nθ) + D
n
sin(nθ) (n ≥ 1)
where C
n
and D
n
are constants.
Now we turn our attention to the R equation: corresponding to the eigenvalue
λ = 0 we obtain the ode r
2
R
//
+ rR
/
= 0. By algebra,
R
//
R
/
= −
1
r
,
which we rewrite as
d
dr
[ln R
/
] = −
1
r
.
Integrating with respect to r,
ln R
/
= −lnr + C,
where C is a constant. Exponentiate both sides and integrate a second time to
obtain
R
0
(r) = E
0
lnr + F
0
,
where E
0
= e
C
and F
0
are constants, and the subscript 0 was introduced to
emphasize that this solution corresponds to the eigenvalue λ = 0.
For the positive eigenvalues λ
n
= n
2
, the ode for R takes the form
r
2
R
//
+ rR
/
−n
2
R = 0. (13.12)
Although this is a linear, secondorder ode, the variable coefﬁcients prevent us
from using a characteristic equation to solve it. Luckily, this is an example of
an Euler equation: a special class of linear differential equations which can be
solved analytically by seeking power function solutions R(r) = r
m
as opposed to
exponential solutions R(r) = e
mr
. For a more careful discussion of how to solve
Eulertype odes, see the Appendix at the end of this section.
In light of these remarks, let us seek solutions of (13.12) of the form R(r) = r
α
,
where α is a real number. The equation becomes
r
2
α(α −1)r
α−2
+ rαr
α−1
−n
2
r
α
= 0,
394 laplace’ s equati on on bounded domai ns
and dividing by r
α
yields
α(α −1) + α −n
2
= 0,
By algebra, a
2
−n
2
= 0, which has solutions α = ±n. It follows that r
−n
and r
n
are solutions of (13.12). Moreover, since these are linearly independent functions
of r and are solutions to a linear, homogeneous ode, the general solution of (13.12)
is
R
n
(r) = E
n
r
−n
+ F
n
r
n
(n ≥ 1),
where E
n
and F
n
are constants. Again, the reason for introducing the subscript n
is to emphasize the correspondence between the functions and the eigenvalues
λ
n
.
Now recall that we are seeking bounded solutions of the Dirichlet prob
lem (13.11). Since the functions lnr and r
−n
blow up as r → 0, we may exclude
these terms in the solutions of the R equation. In other words,
R
n
(r) = F
n
r
n
(n ≥ 0).
The next step is to form the general solution of (13.11) by assembling the
solutions of the separated equations for R and Θ. Deﬁne
u
n
(r, θ) = R
n
(r)Θ
n
(θ) (n ≥ 0)
and form
u(r, θ) =
∞
∑
n=0
G
n
u
n
(r, θ) = G
0
F
0
C
0
+
∞
∑
n=1
G
n
F
n
r
n
[C
n
cos(nθ) + D
n
sin(nθ)] .
Anticipating the use of a Fourier series representation, abbreviate the various
combinations of constants as
A
0
2
= G
0
F
0
C
0
A
n
= G
n
F
n
C
n
and B
n
= G
n
F
n
D
n
(n ≥ 1).
The general solution of (13.11) is
u(r, θ) =
A
0
2
+
∞
∑
n=1
r
n
[A
n
cos(nθ) + B
n
sin(nθ)] . (13.13)
the laplace and poi sson equati ons 395
The last step is to enforce the boundary condition u(a, θ) = h(θ). Setting r = a
in (13.13),
h(θ) =
A
0
2
+
∞
∑
n=1
a
n
[A
n
cos(nθ) + B
n
sin(nθ)] ,
a (full) Fourier series representation of h(θ). Since h(θ) has period 2π, we can
obtain the Fourier coefﬁcients A
n
and B
n
by integrating over any interval of
length 2π. It is convenient to choose the integral [0, 2π] (as opposed to [−π, π])
and to deﬁne
˜
A
n
= a
n
A
n
and
˜
B
n
= a
n
B
n
. Then
˜
A
n
=
1
π
_
2π
0
h(ψ) cos(nψ) dψ (n ≥ 0)
and
˜
B
n
=
1
π
_
2π
0
h(ψ) sin(nψ) dψ (n ≥ 1).
Inserting these Fourier coefﬁcients into formula (13.13), we have shown that the
bounded solution of the Dirichlet problem (13.11) is
u(r, θ) =
1
2π
_
2π
0
h(ψ) dψ
+
1
π
∞
∑
n=1
r
n
a
n
_
2π
0
h(ψ) [cos(nψ) cos(nθ) +sin(nψ) sin(nθ)] dψ.
By the doubleangle identities, the trigonometric terms in the integrand collapse
into a simplerlooking expression:
u(r, θ) =
1
2π
_
2π
0
h(ψ) dψ +
1
π
∞
∑
n=1
r
n
a
n
_
2π
0
h(ψ) cos[n(θ −ψ)] dψ
=
1
2π
_
2π
0
h(ψ)
_
1 + 2
∞
∑
n=1
_
r
a
_
n
cos[n(θ −ψ)]
_
dψ.
(13.14)
Observation. The inﬁnite series appearing in the integrand of (13.14) is actually
a geometric series in disguise, and can be summed explicitly.
To see this, it helps to write the cosine function in terms of complex exponential
functions via Euler’s identity
cos θ =
1
2
_
e
iθ
+e
−iθ
_
.
396 laplace’ s equati on on bounded domai ns
Focusing our attention on the series that appears in Equation (13.14), applying
Euler’s identity to the cosine term yields
1 + 2
∞
∑
n=1
_
r
a
_
n
cos[n(θ −ψ)] = 1 +
∞
∑
n=1
_
r
a
_
n
_
e
in(θ−ψ)
+e
−in(θ−ψ)
_
.
= 1 +
∞
∑
n=1
_
r
a
_
n
e
in(θ−ψ)
. ¸¸ .
Series 1
+
∞
∑
n=1
_
r
a
_
n
e
−in(θ−ψ)
. ¸¸ .
Series 2
.
In the latter expression, Series 1 and Series 2 are both geometric: the ratio
of consecutive terms is independent of n. In Series 1, the common ratio of
consecutive terms is
r
a
e
i(θ−ψ)
,
a complex number with modulus less than one since (r/a) < 1 and the exponen
tial factor has modulus exactly 1. Similar remarks hold for Series 2, and since
the common ratios of these series have moduli less than 1, we can sum these
geometric series explicitly to get
1 +
(r/a)e
i(θ−ψ)
1 −(r/a)e
i(θ−ψ)
. ¸¸ .
sum of Series 1
+
(r/a)e
−i(θ−ψ)
1 −(r/a)e
−i(θ−ψ)
. ¸¸ .
sum of Series 2
= 1 +
re
i(θ−ψ)
a −re
i(θ−ψ)
+
re
−i(θ−ψ)
a −re
−i(θ−ψ)
.
Introducing a common denominator,
1 +
re
i(θ−ψ)
[a −re
−i(θ−ψ)
] + re
−i(θ−ψ)
[a −re
i(θ−ψ)
]
[a −re
i(θ−ψ)
][a −re
−i(θ−ψ)
]
= 1 +
−r
2
+ ar[e
i(θ−ψ)
+e
−i(θ−ψ)
] −r
2
a
2
−2ar[e
i(θ−ψ)
+e
−i(θ−ψ)
] + r
2
.
Finally, Euler’s identity allows us to express this using realvalued functions:
1 +
−2r
2
+2ar cos(θ −ψ)
a
2
−2ar cos(θ −ψ) + r
2
=
a
2
−r
2
a
2
−2ar cos(θ −ψ) + r
2
.
the laplace and poi sson equati ons 397
Hence, we have evaluated the summation that appears in formula (13.14). It
follows that the solution of the Dirichlet problem on a disc (13.11) is given by
u(r, θ) =
a
2
−r
2
2π
_
2π
0
h(ψ)
a
2
−2ar cos(θ −ψ) + r
2
dψ, (13.15)
for 0 ≤ r < a and 0 ≤ θ < 2π. Formula (13.15) is a famous result in pdes, and
is called Poisson’s formula. One reason that Poisson’s formula is so elegant is
that it expresses the solution of our Dirichlet problem as a single integral, as
opposed to the sort of inﬁnite series representation that we have come to expect.
Compare (13.15) to the solutions of Dirichlet problems for the wave and heat
equations on the [onedimensional] domain 0 < x < L (see Equations (11.28)
and (11.34)). For those equations, we were forced to present the solutions as
inﬁnite series, because those series could not be summed explicitly like the one
in (13.14). In that respect, Poisson’s formula is one of the most remarkable (and
lucky) results from basic pdes.
Unfortunately, for most choices of boundary conditions h, the integral (13.15)
is impossible to evaluate by hand. (We encountered similar difﬁculty when we
solved the Cauchy problem for the heat equation.)
Example 13.4.1. Suppose that we wish to solve the Dirichlet problem on a
disc (13.11) with the boundary condition h(θ) = 1; i.e., a constant temperature
distribution along the circular boundary. Poisson’s formula (13.15) states that
the solution is
u(r, θ) =
a
2
−r
2
2π
_
2π
0
1
a
2
−2ar cos(θ −ψ) + r
2
dψ.
Evaluating this integral directly is complicated, which is frustrating given the
apparent simplicity of the boundary condition. Intuitively, we would expect the
temperature within the disc to be a constant function, u(r, θ) = 1 for 0 ≤ r ≤ a
and 0 ≤ θ < 2π.
The problem is much simpler if we retrace our steps in the derivation of (13.15),
instead referring to equation (13.14):
u(r, θ) =
1
2π
_
2π
0
h(ψ) dψ +
1
π
∞
∑
n=1
r
n
a
n
_
2π
0
h(ψ) cos[n(θ −ψ)] dψ
398 laplace’ s equati on on bounded domai ns
for 0 ≤ r < a and 0 ≤ θ < 2π. With h(ψ) = 1, we ﬁnd that
u(r, θ) =
1
2π
_
2π
0
dψ +
1
π
∞
∑
n=1
r
n
a
n
_
2π
0
cos[n(θ −ψ)] dψ.
The integrals are easy to calculate:
_
2π
0
dψ = 2π
and
_
2π
0
cos[n(θ −ψ)] dψ = −
sin[n(θ −ψ)]
n
¸
¸
¸
¸
2π
0
= −
1
n
[sin(nθ −2πn) −sin(nθ)] = 0 (n ≥ 1).
Therefore, all of the integrals appearing in the summation are zero, implying that
the solution of the Dirichlet problem is u(r, θ) = 1 throughout the discshaped
domain (as expected).
Example 13.4.2. Solve the Dirichlet problem
∆u = 0 0 ≤ r < a
u(a, θ) = cos(θ) r = a.
Solution: This time, we are given a nonconstant heat distribution on the circular
boundary of the domain. If we attempt to apply Poisson’s formula directly, we
obtain
u(r, θ) =
a
2
−r
2
2π
_
2π
0
cos(ψ)
a
2
−2ar cos(θ −ψ) + r
2
dψ
for 0 ≤ r < a and 0 ≤ θ < 2π. As in the preceding example, we encounter
an integral that appears difﬁcult to evaluate directly. Instead, we quote (13.14)
where, in this case, h(ψ) = cos ψ:
u(r, θ) =
1
2π
_
2π
0
cos ψ dψ +
1
π
∞
∑
n=1
r
n
a
n
_
2π
0
cos ψ cos[n(θ −ψ)] dψ.
the laplace and poi sson equati ons 399
The leading term is easy to evaluate:
1
2π
_
2π
0
cos ψ dψ = 0.
The integrals appearing in the summation are reminiscent of the ones we en
countered during our study of Fourier series. It helps to apply the trigonometric
identity
cos α cos β =
1
2
[cos(α + β) +cos(α − β)]
to rewrite the integrals as
_
2π
0
cos ψ cos[n(θ −ψ)] dψ =
_
2π
0
cos ψ cos(nθ −nψ) dψ
=
1
2
_
2π
0
cos[nθ −(n −1)ψ)] dψ +
1
2
_
2π
0
cos[−nθ + (n +1)ψ)] dψ.
When calculating antiderivatives of the integrands, we must distinguish between
the case n > 1 and n = 1. If n > 1, then these integrals evaluate to
1
2
sin[nθ −(n −1)ψ)]
−(n −1)
¸
¸
¸
¸
2π
0
+
1
2
sin[−nθ + (n +1)ψ)]
n +1
¸
¸
¸
¸
2π
0
=
sin[nθ −2π(n −1)] −sin(nθ)
−2(n −1)
+
sin[−nθ +2π(n +1)] −sin(−nθ)
2(n +1)
.
Since 2π(n −1) and 2π(n +1) are integer multiples of 2π, this entire expression
reduces to 0.
Finally, if n = 1 we evaluate
1
2
_
2π
0
cos θ dψ +
1
2
_
2π
0
cos(2ψ −θ) dψ = πcos θ,
the only nonzero integral in our inﬁnite series representation of u(r, θ). The
series collapses to a single term:
u(r, θ) =
1
π
_
r
a
_
πcos θ =
_
r
a
_
cos θ,
the solution of our Dirichlet problem for 0 ≤ r ≤ a and 0 ≤ θ < 2π.
Take a moment to interpret this solution. The factor cos θ comes from our
boundary condition, while the factor (r/a) varies from 0 to 1 as we move radially
400 laplace’ s equati on on bounded domai ns
outward from the center of our disc toward the boundary. The temperature at
the center of the disc (r = 0) is zero, which happens to be the average of the
temperature along the circular boundary:
1
2π
_
2π
0
h(ψ) dψ =
1
2π
_
2π
0
cos ψ dψ = 0.
This is no accident, as we now show.
If we set r = 0 in Poisson’s formula, we ﬁnd that the temperature at the center
of the disc is
u(0, θ) =
a
2
2π
_
2π
0
h(ψ)
a
2
dψ =
1
2π
_
2π
0
h(ψ) dψ,
the average value of the temperature along the boundary r = a. This observation
is a special case of a remarkable property of solutions of Laplace’s equation:
Theorem 13.4.3. Mean value property for harmonic functions: Suppose that u
is a harmonic function on an open disc D and is continuous on the closed disc
D∪∂D. Then the value of u at the center of the disc is equal to the average value
of u on the boundary ∂D.
Example 13.4.4. Suppose that u is a solution of the Dirichlet problem (13.11)
with boundary condition h(θ) = 10 + θ sin θ. Then the value of u at the origin is
given by
1
2π
_
2π
0
10 + θ sin θ dθ = 10 +
1
2π
_
2π
0
θ sin θ dθ.
= 10 +
−θ cos θ +sin θ
2π
¸
¸
¸
¸
2π
0
= 9.
Finding the maximum and minimum temperatures on this disc is tricky. By
the Maximum Principle 13.2.1, we know that the maximum and minimum
temperatures must be achieved on the boundary. Thus, we must ﬁnd the extreme
values of h(θ) = 10 + θ sin θ on the interval 0 ≤ θ ≤ 2π. Since h
/
(θ) = sin θ +
θ cos θ, critical points occur when θ = −tan θ. Unfortunately, it is impossible
to solve this transcendental equation by hand, and we would need to resort
to computer assistance. It turns out that there are two critical points in the
open interval θ ∈ (0, 2π), one of which corresponds to the global minimum
temperature, and one of which corresponds to the global maximum temperature.
the laplace and poi sson equati ons 401
Additional properties of harmonic functions. Poisson’s formula (13.15) and
the mean value property (13.4.3) lead to a host of important theorems that
characterize the behavior of harmonic functions. We state a few such results for
reference—see the texts of McOwen [6] and Strauss [10] for technical proofs.
Smoothness Theorem: Suppose u is a harmonic function on an open set Ω. Then
u is smooth (i.e., has derivatives of all orders) on Ω. The Smoothness Theorem may
seem surprising given that Laplace’s equation ∆u = 0 only requires existence of
the second partial derivatives of u. In this sense, the class of harmonic functions
is far more “selective” than we might expect.
Harnack’s Theorem: Suppose that u is a nonnegative harmonic function on an
open set Ω, and let Ω
1
be a closed subset of Ω. Then there exists a constant C depending
only on Ω
1
such that
max
Ω
1
u ≤ Cmin
Ω
1
u.
The main implication of Harnack’s Theorem is that nonnegative harmonic
functions cannot exhibit crazy oscillations on closed sets.
Liouville’s Theorem: Suppose that u is a harmonic function in the whole plane
R
2
(or in the whole space R
3
). If u is bounded, then u is a constant function.
Appendix: Solving Euler equations. Consider the linear, homogeneous ode
t
2
u
//
(t) + βtu
/
(t) + γu(t) = 0,
where β and γ are constants. Notice that the exponents of the coefﬁcient functions
involve powers of the independent variable: t
2
, t
1
and t
0
. For this reason, it
is natural to seek power function solutions of the form u(t) = t
m
, because
differentiating such functions reduces the exponent by one (unless m = 0).
Inserting u = t
m
into the above equation yields
t
2
m(m−1)t
m−2
+ βtmt
m−1
+ γt
m
= 0,
which simpliﬁes to
t
m
[m(m−1) + βm + γ] = 0.
Since the left hand side is identically equal to zero as a function of t, it must be
the case that
m(m−1) + βm + γ = 0.
402 laplace’ s equati on on bounded domai ns
This quadratic equation for m is called an indicial equation for our Euler equation,
and is analogous to a characteristic equation for a constantcoefﬁcient ode. If the
indicial equation has distinct, real roots m = r and m = s, then the functions t
r
and t
s
are linearly independent solutions of the original ode. By linearity, the
general solution would then be given by
u(t) = C
1
t
r
+ C
2
t
s
,
where C
1
and C
2
are arbitrary constants.
Exercises
1. Consider the heat equation u
t
= κu
xx
on the interval 0 ≤ x ≤ L with
Dirichlet boundary conditions u(0, t) = τ
1
and u(L, t) = τ
2
. Find the steady
state solution of this Dirichlet problem, and give a mathematical/physical
interpretation of the result.
2. Solve u
xx
+ u
yy
+ u
zz
= 0 in the spherical shell 1 < r < 4 with the Dirichlet
boundary conditions u = τ
1
when r = 1 and u = τ
2
when r = 4. Here, τ
1
and
τ
2
are constants, and r =
_
x
2
+ y
2
+ z
2
measures distance from the origin.
3. Find the bounded solution of u
xx
+ u
yy
= 3
_
x
2
+ y
2
in the disc 0 ≤ r < 3 with
u = 4 on the boundary r = 3. As usual, r =
_
x
2
+ y
2
.
4. Solve u
xx
+ u
yy
+ u
zz
= 6 in the spherical shell 2 < r < 3 with u(x, y, z) = 4
on both the inner and outer shells of the boundary.
5. Solve u
xx
+ u
yy
+ u
zz
= 20(x
2
+ y
2
+ z
2
) in the spherical shell a < r < b with
the Dirichlet condition u = 0 on the inner boundary r = a and the Neumann
condition ∂u/∂r = 0 on the outer boundary r = b.
6. Solve u
xx
+ u
yy
= 0 inside the rectangular domain 0 < x < a and 0 < y < b
with Dirichlet boundary conditions u = 0 on the top, bottom, and left edges
of the rectangle, but with u(a, y) = R(y) on the right edge.
7. Solve u
xx
+ u
yy
= 0 inside the square domain 0 < x < 1 and 0 < y < 1
with boundary conditions u = 0 on the bottom, left, and right edges of the
rectangle, but with the Neumann condition u
y
(x, 1) = T(x) on the top edge.
8. Solve u
xx
+ u
yy
= 0 inside the square domain 0 < x < 2 and 0 < y < 2 with
homogeneous Dirichlet boundary conditions u = 0 on the bottom, left, and
top edges of the rectangle, and with the inhomogeneous Dirichlet condition
the laplace and poi sson equati ons 403
u(2, y) = 1 on the right edge. Then, repeat the calculation with u(2, y) = 1
replaced by the inhomogeneous Neumann condition u
x
(2, y) = 1 on the right
edge.
9. Suppose D is the open disc r < 7 (centered at the origin),
¯
D is the correspond
ing closed disc r ≤ 7, and ∂D denotes the boundary r = 7. Suppose that u is
a harmonic function inside D and that u = 50 + 10 sin2θ + 10 cos 2θ on ∂D.
Without solving for u, determine
(a) the maximum value of u in
¯
D, and
(b) the value of u at the origin.
10. Solve u
xx
+ u
yy
= 0 in the disc r < a with the boundary condition u =
8 −5 sin θ when r = a. Hint: You may ﬁnd it difﬁcult to use Poisson’s formula
directly, so refer to (13.14) instead.
Guide to Commonly Used Notation
Symbol Usual Meaning
R the set of real numbers
R
n
ndimensional Euclidean space
t independent variable (time)
u, v vectors consisting of real numbers
x, y vectors consisting of dependent variables
x
0
, y
0
vectors consisting of initial conditions x(0), y(0)
x
∗
, y
∗
equilibrium solution of a system of odes
f a vector consisting of functions
A, M, P square n n matrices
D a diagonal n n matrix
N a nilpotent n n matrix
tr(A) trace of a square matrix A
det(A) determinant of a square matrix A
λ eigenvalue of a matrix, or a Lyapunov exponent
α, β real, imaginary parts of an eigenvalue λ = α + βi
E
s
, E
u
, E
c
stable, unstable, and center subspaces
W
s
, W
u
, W
c
stable, unstable, and center manifolds
A ⊕B direct sum of subspaces A and B
span{v
1
, v
2
, . . . v
n
} span of vectors v
1
, v
2
, . . . v
n
φ
t
(x
0
) solution of x
/
= f (x) with x(0) = x
0
φ
t
ﬂow of a system of odes
f : R
n
→R
m
a function from R
n
into R
m
∇f gradient of a function from R
n
into R
J f Jacobian matrix of a function from R
n
into R
m
u • v dot product of vectors u and v
v
2
Euclidean norm (length) of a vector v in R
n
u −v
2
Euclidean distance between points u and v in R
n
B(x, ) open ball of radius centered at x
V(x) a Lyapunov function
Γ(t) a parametrized curve, Γ : R →R
n
404
gui de to commonly used notati on 405
Symbol Usual Meaning
n a normal vector to a parametrized curve
r, θ radius and angle (polar coordinates)
µ a bifurcation parameter
σ Lyapunov number (associated with Hopf bifurcations)
τ time delay (for delay differential equations)
x(a
+
) the righthand limit lim
t→a
+ x(t)
x(a
−
) the lefthand limit lim
t→a
− x(t)
x
∗
a ﬁxedpoint of a ﬁrstorder difference equation
Sf (x) Schwarzian derivative of a function f : R →R
γ feedback gain parameter in the tdas algorithm
L linear operator, or length of the interval [0, L]
φ(x) initial condition for a pde
∂Ω boundary of the domain Ω
κ positivevalued diffusion constant (heat equation)
c wave speed (wave/transport equations)
ψ(x) initial velocity (wave equation)
S φ convolution of functions S and φ
S(x, t) (onedimensional) heat kernel
φ
odd
, φ
even
odd, even extensions of a function φ
X(x), Y(y), T(t) separated solutions of pdes (separation of variables)
R(r), Θ(θ) separated solutions of pdes (polar coordinates)
β abbreviation for square roots of eigenvalues λ
A
n
, B
n
Fourier coefﬁcients
f 
∞
, f 
L
2 different types of norms of the function f
¸ f , g¸ inner product of functions f and g
S
N
(x) partial sum of an inﬁnite series of functions
f (x
−
0
), f (x
+
0
) left and righthand limits of f (x) as x → x
0
f
/
(x
−
0
), f
/
(x
+
0
) left and righthand derivatives of f (x) as x → x
0
∆ Laplace operator
v(r) a radial solution of Laplace’s or Poisson’s equation
h(θ) boundary condition for Laplace’s equation on a disc
References
[1] G. Bachman, L. Narici, and E. Beckenstein, Fourier and wavelet analysis, Springer
Verlag, New York, 2000.
[2] R. Bellman and K. L. Cooke, Differentialdifference equations, Academic Press, New
York, 1963.
[3] S. Elaydi, An introduction to difference equations, Springer, New York, 1999.
[4] B.D. Hassard, N.D. Kazarinoff, and Y.H. Wan, Theory and applications of hopf bifurca
tion, Cambridge University Press, New York, 1981.
[5] J. H. Hubbard and B. H. West, Differential equations: A dynamical systems approach,
SpringerVerlag, New York, 1997.
[6] R. C. McOwen, Partial differential equations: Methods and applications, Prentice Hall,
Upper Saddle River, 2003.
[7] J. D. Meiss, Differential dynamical systems, SIAM, Philadelphia, 2007.
[8] L. Perko, Differential equations and dynamical systems, SpringerVerlag, New York, 2001.
[9] J. E. S. Socolar, D. W. Sukow, and D. J. Gauthier, Stabilizing unstable periodic orbits in
fast dynamical systems, Physical Review E 50 (1994), 3245–3248.
[10] W. A. Strauss, Partial differential equations: An introduction, 2nd ed., Wiley, New York,
2007.
[11] S. Strogatz, Nonlinear dynamics and chaos, AddisonWesley, 1994.
406
Index
bvp, see boundary value problem
dde, see delay differential equation
etdas, 212
fft, see fast Fourier transform
ic, see initial condition
ivp, see initial value problem
ode, see ordinary differential equation
pde, see partial differential equation
tdas, see timedelay autosynchronization
2cycle, 202
advection equation, 219, 239
advectiondiffusion equation, 294
algebraic multiplicity, 21
AndonovHopf bifurcation, 152
annular domain, 380
aperiodic, 129
asymptotically stable equilibrium, 57
asymptotically stable ﬁxed point, 193
attractor, 109, 193
autonomous, 78
basin of attraction, 161
bifurcation, 140
bifurcation diagram, 141, 203
bistability, 163
block diagonal matrix, 35
blow up, 81
boundary, 370
boundary condition, 5, 231
boundary value problem, 5
bounded set, 371
Burgers’ equation, 219, 246
catastrophe, 151
Cauchy problem, 230
center, 52
center subspace, 58
chaos, 129, 204
characteristic curves, 238
characteristic equation, 17, 171, 181
characteristic polynomial, 17
closed formula, 181
complex conjugate, 29
connected set, 371
continuously differentiable, 87, 284
control domain, 211
convergence in the L
2
sense (sequences), 354
convergence in the L
2
sense (series), 356
convolution, 266
D’Alembert’s formula, 259
degenerate equilibrium, 52
delay differential equation, 166
dependence on initial conditions, 89
deterministic, 129
diagonal matrix, 12
diagonalizable matrix, 14
diagonalization, 19
difference equation, 180
diffusion equation, 219
dilation, 266
direct sum of subspaces, 64
directional derivative, 236
Dirichlet boundary condition, 231
Dirichlet problem on a disc, 390
discrete logistic equation, 200
discrete system, 180
domain of dependence, 264
domain of inﬂuence, 264
dot product, 124
Duhamel’s Principle, 284
eigenfunction, 316
eigenspace, 21
eigenvalue, 15, 316
eigenvector, 15, 316
electrostatic potential, 370
elliptic pde, 226
energy method, 276
equilibrium, 48, 94
Euclidean distance, 89
Euclidean norm, 89
Euler equation, 393, 401
407
408
even extension, 302, 346
even function, 292, 345
existence, 90
existence and uniqueness, see Fundamental Ex
istence & Uniqueness Theorem
exponential of a matrix, see matrix exponential
extended tdas, 212
extreme value theorem, 87
fast Fourier transform, 205
Feigenbaum constant, 203
Fibonacci sequence, 181, 185
Fick’s law of diffusion, 251
FitzHughNagumo model, 137
ﬁxed point, 191, 196
ﬂow, 63, 101
focus, 52
Fourier cosine series, 323, 331
Fourier cosine series coefﬁcients, 339
Fourier series, 327, 331
Fourier series coefﬁcients, 343
Fourier sine series, 313, 331
Fourier sine series: coefﬁcients, 334
full Fourier series, see Fourier series, see Fourier
series
functional analysis, 316
Fundamental Existence & Uniqueness Theorem,
85
Gaussian function, 272
generalized eigenvector, 39
geometric multiplicity, 21
geometric series, 356
global existence, 87, 92
globally asymptotically stable, 112
glycolysis, 164
gradient, 82
gradient system, 113
Gronwall’s inequality, 88, 115
Gronwall’s inequality: strong, 92
harmonic function, 369
Harnack’s theorem, 401
HartmanGrobman theorem, 99
heat equation, 219, 250, 286
heat equation: Cauchy problem, 265
heat equation: Dirichlet problem, 313
heat equation: Dirichlet problem on a halfline,
297
heat equation: Neumann problem, 324
heat equation: Neumann problem on a halfline,
301
heat kernel, 272
heteroclinic orbit, 130, 132
homoclinic orbit, 134
homogeneous, 10
homogeneous pde, 224
Hopf bifurcation, see AndronovHopf bifurca
tion
Hopf bifurcation theorem, 158, 159
HutchinsonWright equation, 173
hyperbolic pde, 226
hyperbolic equilibrium, 63, 98
hyperbolic ﬁxed point, 194
hyperbolic ﬂow, 63
hyperbolic system, 63
hysteresis, 163
idempotent matrix, 72
imaginary part of a complex number, 29
indicial equation, 402
inhomogeneous, 10
inhomogeneous pde, 224, 284
inhomogeneous linear system, 67
initial condition, 5, 229
initial function, 168
initial value problem, 5, 230
inner product, 351
integrating factor, 68
invariant, 101
isolated equilibrium, 94
jacobian, 83
Jordan decomposition, 39
Jordan form, 39
jump discontinuity, 361
Jury stability test, 199
KortewegdeVries equation, 220
L
2
convergence, see convergence in the L
2
sense
(sequences)
409
L
2
distance, 279, 349
L
2
inner product, 350
L
2
norm, 348
L
∞
distance, 349
Laplace equation, 220, 369
Laplace operator, 369
Laplace’s equation on a rectangular domain, 384
Laplace’s equation: Dirichlet problem, 371
Laplace’s equation: fundamental solution in 2d,
378
Laplace’s equation: fundamental solution in 3d,
380
Laplace’s equation: Neumann problem, 372
Laplacian, see Laplace operator
lefthand derivative, 361
lefthand limit, 361
limit cycle, 128
linear, 79
linear pde, 223
linear approximation, 81
linear combination, 225
linear operator, 221
linearization, 96
Liouville’s theorem, 401
locally asymptotically stable, 109
locally stable, 108
logistic equation, 172
logistic equation: discrete, 200
Lorenz equations, 129
LotkaVolterra model, 118
lower triangular, 24
Lyapunov exponent, 207
Lyapunov function, 111
Lyapunov number, 158
Lyapunov theorem, 110
manifold, 100
mapping, kdimensional, 180
matrix exponential, 11
maximal interval of existence, 91
maximum principle: heat equation, 280
maximum principle: Laplace’s equation, 372
mean value property of harmonic functions, 400
mean value theorem, 87
method of characteristics, 238
method of even extension, 302
method of odd extension, 297
method of steps, 168
minimum principle: heat equation, 281
Minkowski’s inequality, 348
mixed boundary conditions, 324
modulus of a complex number, 197
Neumann boundary condition, 232
neutral dde, 167
Newton’s second law of motion, 252
nilpotent matrix, 39
node, 49
nondiagonalizable matrix, 38
nonhyperbolic ﬁxed point, 194
nonlinear, 79
nonlinear operator, 222
norm, 348
normal derivative, 232, 372
normal vector, 123
nullclines, 120
odd extension, 298, 346
odd function, 298, 345
open ball, 94
open set, 370
operator, 221
order, 3, 219
ordinary differential equation, 2
parabolic pde, 226
parameter, 140
partial differential equation, 2, 219
partial sum, 356
period2 solution, 202
perioddoubling bifurcation, 201
periodic boundary conditions, 329, 367
periodic extension, 346
periodic function, 345
periodic orbit, 122
periodic solution, 122
phase plane, 46
phase portrait, 46
Picard’s method of successive approximations,
84
piecewise smooth, 370
410
pitchfork bifurcation, 143
planar system, 45
PoincaréBendixon theorem, 126
pointwise convergence (sequences), 353
pointwise convergence (series), 356
Poisson equation, 370
Poisson’s formula, 397
potential function, 114
predatorprey model, 118
propagation speed, 274
reactiondiffusion equation, 294, 316
real canonical form, 32
real part of a complex number, 29
reﬂection method, 297
repeller, 108, 193
retarded dde, 167
righthand derivative, 361
righthand limit, 361
rotation invariance, 375
saddle, 49
saddlenode bifurcation, 141
Schwarzian derivative, 195
second iterate mapping, 201
separation of variables, 307
separatrix, 50
shock wave, 247
similar matrices, 14
simple, closed curve, 370
sink, 109
smoothness theorem, 401
solution of a pde, 220
solution of an ode, 3
span, 58
stable equilibrium, 57
stable ﬁxed point, 193
stable focus, 52
stable manifold, 102
stable manifold theorem, 102
stable node, 49
stable periodic orbit, 127
stable subspace, 58
steadystate, 368
strong Gronwall inequality, see Gronwall’s in
equality: strong
subcritical bifurcation, 145
sum of two subspaces, 64
supercritical bifurcation, 145
superposition principle, 225
system of odes, 2
tangent plane, 81
Taylor series, 330
timedelay autosynchronization, 208
trace of a matrix, 65
transcendental equation, 172
transcritical bifurcation, 142
translation invariance, 265, 375
transport equation, 219, 239, 291
transversality condition, 159, 176
triangle inequality, 89
triangular matrix, 24
uncoupled system, 13
uniform convergence (sequences), 351
uniform convergence (series), 356
uniqueness, 90
unit normal vector, 123
unstable, 109
unstable equilibrium, 57
unstable ﬁxed point, 193
unstable focus, 52
unstable manifold, 102
unstable node, 49
unstable periodic orbit, 128
unstable subspace, 58
upper triangular, 24
variation of parameters, 68, 284
wave equation, 219, 252, 288
wave equation: Cauchy problem, 257
wave equation: Dirichlet problem, 307
wave equation: Dirichlet problem on a halfline,
303
wave equation: Neumann problem, 318
wave equation: Neumann problem on a half
line, 306
wellposed, 90, 232, 276, 375